Create container image and move parakeet fork to vendor dir

16 files changed, 3396 insertions, 0 deletions

diff --git a/vendor/parakeet-rs/src/audio.rs b/vendor/parakeet-rs/src/audio.rs
new file mode 100644
index 0000000..84d2616
--- /dev/null
+++ b/vendor/parakeet-rs/src/audio.rs
@@ -0,0 +1,179 @@
+use crate::config::PreprocessorConfig;
+use crate::error::{Error, Result};
+use hound::{WavReader, WavSpec};
+use ndarray::Array2;
+use std::f32::consts::PI;
+use std::path::Path;
+
+pub fn load_audio<P: AsRef<Path>>(path: P) -> Result<(Vec<f32>, WavSpec)> {
+    let mut reader = WavReader::open(path)?;
+    let spec = reader.spec();
+
+    let samples: Vec<f32> = match spec.sample_format {
+        hound::SampleFormat::Float => reader
+            .samples::<f32>()
+            .collect::<std::result::Result<Vec<_>, _>>()
+            .map_err(|e| Error::Audio(format!("Failed to read float samples: {e}")))?,
+        hound::SampleFormat::Int => reader
+            .samples::<i16>()
+            .map(|s| s.map(|s| s as f32 / 32768.0))
+            .collect::<std::result::Result<Vec<_>, _>>()
+            .map_err(|e| Error::Audio(format!("Failed to read int samples: {e}")))?,
+    };
+
+    Ok((samples, spec))
+}
+
+pub fn apply_preemphasis(audio: &[f32], coef: f32) -> Vec<f32> {
+    let mut result = Vec::with_capacity(audio.len());
+    result.push(audio[0]);
+
+    for i in 1..audio.len() {
+        result.push(audio[i] - coef * audio[i - 1]);
+    }
+
+    result
+}
+
+fn hann_window(window_length: usize) -> Vec<f32> {
+    (0..window_length)
+        .map(|i| 0.5 - 0.5 * ((2.0 * PI * i as f32) / (window_length as f32 - 1.0)).cos())
+        .collect()
+}
+
+// We use proper FFT here instead of naive DFT because the model was trained
+// on correctly computed spectrograms. Naive DFT produces wrong frequency bins
+// and the model outputs all blank tokens. RustFFT gives us O(n log n) performance
+// and numerically correct results that match what the model expects.
+pub fn stft(audio: &[f32], n_fft: usize, hop_length: usize, win_length: usize) -> Array2<f32> {
+    use rustfft::{num_complex::Complex, FftPlanner};
+
+    let window = hann_window(win_length);
+    let num_frames = (audio.len() - win_length) / hop_length + 1;
+    let freq_bins = n_fft / 2 + 1;
+    let mut spectrogram = Array2::<f32>::zeros((freq_bins, num_frames));
+
+    let mut planner = FftPlanner::<f32>::new();
+    let fft = planner.plan_fft_forward(n_fft);
+
+    for frame_idx in 0..num_frames {
+        let start = frame_idx * hop_length;
+
+        let mut frame: Vec<Complex<f32>> = vec![Complex::new(0.0, 0.0); n_fft];
+        for i in 0..win_length.min(audio.len() - start) {
+            frame[i] = Complex::new(audio[start + i] * window[i], 0.0);
+        }
+
+        fft.process(&mut frame);
+
+        for k in 0..freq_bins {
+            let magnitude = frame[k].norm();
+            spectrogram[[k, frame_idx]] = magnitude * magnitude;
+        }
+    }
+
+    spectrogram
+}
+
+fn hz_to_mel(freq: f32) -> f32 {
+    2595.0 * (1.0 + freq / 700.0).log10()
+}
+
+fn mel_to_hz(mel: f32) -> f32 {
+    700.0 * (10.0_f32.powf(mel / 2595.0) - 1.0)
+}
+
+fn create_mel_filterbank(n_fft: usize, n_mels: usize, sample_rate: usize) -> Array2<f32> {
+    let freq_bins = n_fft / 2 + 1;
+    let mut filterbank = Array2::<f32>::zeros((n_mels, freq_bins));
+
+    let min_mel = hz_to_mel(0.0);
+    let max_mel = hz_to_mel(sample_rate as f32 / 2.0);
+
+    let mel_points: Vec<f32> = (0..=n_mels + 1)
+        .map(|i| mel_to_hz(min_mel + (max_mel - min_mel) * i as f32 / (n_mels + 1) as f32))
+        .collect();
+
+    let freq_bin_width = sample_rate as f32 / n_fft as f32;
+
+    for mel_idx in 0..n_mels {
+        let left = mel_points[mel_idx];
+        let center = mel_points[mel_idx + 1];
+        let right = mel_points[mel_idx + 2];
+
+        for freq_idx in 0..freq_bins {
+            let freq = freq_idx as f32 * freq_bin_width;
+
+            if freq >= left && freq <= center {
+                filterbank[[mel_idx, freq_idx]] = (freq - left) / (center - left);
+            } else if freq > center && freq <= right {
+                filterbank[[mel_idx, freq_idx]] = (right - freq) / (right - center);
+            }
+        }
+    }
+
+    filterbank
+}
+
+/// Extract mel spectrogram features from raw audio samples.
+///
+/// # Arguments
+///
+/// * `audio` - Audio samples as f32 values
+/// * `sample_rate` - Sample rate in Hz
+/// * `channels` - Number of audio channels
+/// * `config` - Preprocessor configuration
+///
+/// # Returns
+///
+/// 2D array of mel spectrogram features (time_steps x feature_size)
+pub fn extract_features_raw(
+    mut audio: Vec<f32>,
+    sample_rate: u32,
+    channels: u16,
+    config: &PreprocessorConfig,
+) -> Result<Array2<f32>> {
+    if sample_rate != config.sampling_rate as u32 {
+        return Err(Error::Audio(format!(
+            "Audio sample rate {} doesn't match expected {}. Please resample your audio first.",
+            sample_rate, config.sampling_rate
+        )));
+    }
+
+    if channels > 1 {
+        let mono: Vec<f32> = audio
+            .chunks(channels as usize)
+            .map(|chunk| chunk.iter().sum::<f32>() / channels as f32)
+            .collect();
+        audio = mono;
+    }
+
+    audio = apply_preemphasis(&audio, config.preemphasis);
+
+    let spectrogram = stft(&audio, config.n_fft, config.hop_length, config.win_length);
+
+    let mel_filterbank =
+        create_mel_filterbank(config.n_fft, config.feature_size, config.sampling_rate);
+    let mel_spectrogram = mel_filterbank.dot(&spectrogram);
+    let mel_spectrogram = mel_spectrogram.mapv(|x| (x.max(1e-10)).ln());
+
+    let mut mel_spectrogram = mel_spectrogram.t().to_owned();
+
+    // Normalize each feature dimension to mean=0, std=1
+    let num_frames = mel_spectrogram.shape()[0];
+    let num_features = mel_spectrogram.shape()[1];
+
+    for feat_idx in 0..num_features {
+        let mut column = mel_spectrogram.column_mut(feat_idx);
+        let mean: f32 = column.iter().sum::<f32>() / num_frames as f32;
+        let variance: f32 =
+            column.iter().map(|&x| (x - mean).powi(2)).sum::<f32>() / num_frames as f32;
+        let std = variance.sqrt().max(1e-10);
+
+        for val in column.iter_mut() {
+            *val = (*val - mean) / std;
+        }
+    }
+
+    Ok(mel_spectrogram)
+}
diff --git a/vendor/parakeet-rs/src/config.rs b/vendor/parakeet-rs/src/config.rs
new file mode 100644
index 0000000..1dae890
--- /dev/null
+++ b/vendor/parakeet-rs/src/config.rs
@@ -0,0 +1,51 @@
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct PreprocessorConfig {
+    pub feature_extractor_type: String,
+    pub feature_size: usize,
+    pub hop_length: usize,
+    pub n_fft: usize,
+    pub padding_side: String,
+    pub padding_value: f32,
+    pub preemphasis: f32,
+    pub processor_class: String,
+    pub return_attention_mask: bool,
+    pub sampling_rate: usize,
+    pub win_length: usize,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ModelConfig {
+    pub architectures: Vec<String>,
+    pub vocab_size: usize,
+    pub pad_token_id: usize,
+}
+
+impl Default for PreprocessorConfig {
+    fn default() -> Self {
+        Self {
+            feature_extractor_type: "ParakeetFeatureExtractor".to_string(),
+            feature_size: 80,
+            hop_length: 160,
+            n_fft: 512,
+            padding_side: "right".to_string(),
+            padding_value: 0.0,
+            preemphasis: 0.97,
+            processor_class: "ParakeetProcessor".to_string(),
+            return_attention_mask: true,
+            sampling_rate: 16000,
+            win_length: 400,
+        }
+    }
+}
+
+impl Default for ModelConfig {
+    fn default() -> Self {
+        Self {
+            architectures: vec!["ParakeetForCTC".to_string()],
+            vocab_size: 1025,
+            pad_token_id: 1024,
+        }
+    }
+}
diff --git a/vendor/parakeet-rs/src/decoder.rs b/vendor/parakeet-rs/src/decoder.rs
new file mode 100644
index 0000000..6da6d65
--- /dev/null
+++ b/vendor/parakeet-rs/src/decoder.rs
@@ -0,0 +1,211 @@
+use crate::error::{Error, Result};
+use ndarray::Array2;
+use std::path::Path;
+
+// Token with its timestamp information
+// start and end are in seconds
+#[derive(Debug, Clone)]
+pub struct TimedToken {
+    pub text: String,
+    pub start: f32,
+    pub end: f32,
+}
+
+#[derive(Debug, Clone)]
+pub struct TranscriptionResult {
+    pub text: String,
+    pub tokens: Vec<TimedToken>,
+}
+
+// CTC decoder for parakeet-ctc-0.6b model with token-level timestamps
+pub struct ParakeetDecoder {
+    tokenizer: tokenizers::Tokenizer,
+    pad_token_id: usize,
+}
+
+impl ParakeetDecoder {
+    pub fn from_pretrained<P: AsRef<Path>>(tokenizer_path: P) -> Result<Self> {
+        let tokenizer_path = tokenizer_path.as_ref();
+
+        let tokenizer = tokenizers::Tokenizer::from_file(tokenizer_path)
+            .map_err(|e| Error::Tokenizer(format!("Failed to load tokenizer: {e}")))?;
+
+        // Hardcoded pad_token_id for Parakeet-CTC-0.6b (constant across all models: please see def configs jsons: https://huggingface.co/onnx-community/parakeet-ctc-0.6b-ONNX/tree/main)
+        let pad_token_id = 1024;
+
+        Ok(Self {
+            tokenizer,
+            pad_token_id,
+        })
+    }
+
+    pub fn decode(&self, logits: &Array2<f32>) -> Result<String> {
+        let time_steps = logits.shape()[0];
+
+        let mut token_ids = Vec::new();
+        for t in 0..time_steps {
+            let logits_t = logits.row(t);
+            let max_idx = logits_t
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+                .map(|(idx, _)| idx)
+                .unwrap_or(0);
+
+            token_ids.push(max_idx as u32);
+        }
+
+        let collapsed = self.ctc_collapse(&token_ids);
+
+        let text = self
+            .tokenizer
+            .decode(&collapsed, true)
+            .map_err(|e| Error::Tokenizer(format!("Failed to decode: {e}")))?;
+
+        Ok(text)
+    }
+
+    fn ctc_collapse(&self, token_ids: &[u32]) -> Vec<u32> {
+        let mut result = Vec::new();
+        let mut prev_token: Option<u32> = None;
+
+        for &token_id in token_ids {
+            if token_id == self.pad_token_id as u32 {
+                prev_token = Some(token_id);
+                continue;
+            }
+
+            if Some(token_id) != prev_token {
+                result.push(token_id);
+            }
+
+            prev_token = Some(token_id);
+        }
+
+        result
+    }
+
+    // CTC collapse with frame tracking for timestamps
+    fn ctc_collapse_with_frames(&self, token_ids: &[(u32, usize)]) -> Vec<(u32, usize, usize)> {
+        let mut result: Vec<(u32, usize, usize)> = Vec::new();
+        let mut prev_token: Option<u32> = None;
+
+        for &(token_id, frame) in token_ids.iter() {
+            if token_id == self.pad_token_id as u32 {
+                prev_token = Some(token_id);
+                continue;
+            }
+
+            if Some(token_id) != prev_token {
+                if let Some(prev) = prev_token {
+                    if prev != self.pad_token_id as u32 {
+                        // End previous token
+                        if let Some(last) = result.last_mut() {
+                            last.2 = frame;
+                        }
+                    }
+                }
+                // Start new token
+                result.push((token_id, frame, frame));
+            }
+
+            prev_token = Some(token_id);
+        }
+
+        // Close last token
+        if let Some(last) = result.last_mut() {
+            last.2 = token_ids.len();
+        }
+
+        result
+    }
+
+    // Decode with token-level timestamps
+    // hop_length and sample_rate are needed to convert frames to seconds
+    pub fn decode_with_timestamps(
+        &self,
+        logits: &Array2<f32>,
+        hop_length: usize,
+        sample_rate: usize,
+    ) -> Result<TranscriptionResult> {
+        let time_steps = logits.shape()[0];
+
+        let mut token_ids_with_frames = Vec::new();
+        for t in 0..time_steps {
+            let logits_t = logits.row(t);
+            let max_idx = logits_t
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+                .map(|(idx, _)| idx)
+                .unwrap_or(0);
+
+            token_ids_with_frames.push((max_idx as u32, t));
+        }
+
+        // CTC collapse with frame tracking
+        let collapsed_with_frames = self.ctc_collapse_with_frames(&token_ids_with_frames);
+
+        // Extract just token IDs for decoding
+        let token_ids: Vec<u32> = collapsed_with_frames.iter().map(|(id, _, _)| *id).collect();
+
+        // Decode full text
+        let full_text = self
+            .tokenizer
+            .decode(&token_ids, true)
+            .map_err(|e| Error::Tokenizer(format!("Failed to decode: {e}")))?;
+
+        // Progressive decode to detect word boundaries
+        // BPE tokenizers only add spaces when decoding sequences, not individual tokens
+        let mut timed_tokens = Vec::new();
+        let mut prev_decode = String::new();
+
+        for (i, (_token_id, start_frame, end_frame)) in collapsed_with_frames.iter().enumerate() {
+            // Decode from start up to and including current token
+            let token_ids_so_far: Vec<u32> = collapsed_with_frames[0..=i]
+                .iter()
+                .map(|(id, _, _)| *id)
+                .collect();
+
+            if let Ok(curr_decode) = self.tokenizer.decode(&token_ids_so_far, true) {
+                // Find what this token added
+                let added_text = if curr_decode.len() > prev_decode.len() {
+                    &curr_decode[prev_decode.len()..]
+                } else {
+                    ""
+                };
+
+                if !added_text.is_empty() {
+                    let start_time = (*start_frame * hop_length) as f32 / sample_rate as f32;
+                    let end_time = (*end_frame * hop_length) as f32 / sample_rate as f32;
+
+                    timed_tokens.push(TimedToken {
+                        text: added_text.to_string(),
+                        start: start_time,
+                        end: end_time,
+                    });
+                }
+
+                prev_decode = curr_decode;
+            }
+        }
+
+        Ok(TranscriptionResult {
+            text: full_text,
+            tokens: timed_tokens,
+        })
+    }
+
+    // Stub - falls back to greedy decoding. Full beam search with language model is TODO.
+    pub fn decode_with_beam_search(
+        &self,
+        logits: &Array2<f32>,
+        _beam_width: usize,
+    ) -> Result<String> {
+        self.decode(logits)
+    }
+
+    pub fn pad_token_id(&self) -> usize {
+        self.pad_token_id
+    }
+}
diff --git a/vendor/parakeet-rs/src/decoder_tdt.rs b/vendor/parakeet-rs/src/decoder_tdt.rs
new file mode 100644
index 0000000..65f576d
--- /dev/null
+++ b/vendor/parakeet-rs/src/decoder_tdt.rs
@@ -0,0 +1,63 @@
+use crate::decoder::TranscriptionResult;
+use crate::error::Result;
+use crate::vocab::Vocabulary;
+
+/// TDT greedy decoder for Parakeet TDT models
+#[derive(Debug)]
+pub struct ParakeetTDTDecoder {
+    vocab: Vocabulary,
+}
+
+impl ParakeetTDTDecoder {
+    /// Load decoder from vocab file
+    pub fn from_vocab(vocab: Vocabulary) -> Self {
+        Self { vocab }
+    }
+
+    /// Decode tokens with timestamps
+    /// For TDT models, greedy decoding is done in the model, here we just convert to text
+    pub fn decode_with_timestamps(
+        &self,
+        tokens: &[usize],
+        frame_indices: &[usize],
+        _durations: &[usize],
+        hop_length: usize,
+        sample_rate: usize,
+    ) -> Result<TranscriptionResult> {
+        let mut result_tokens = Vec::new();
+        let mut full_text = String::new();
+         // TDT encoder does 8x subsampling
+        let encoder_stride = 8;
+
+        for (i, &token_id) in tokens.iter().enumerate() {
+            if let Some(token_text) = self.vocab.id_to_text(token_id) {
+                let frame = frame_indices[i];
+                let start = (frame * encoder_stride * hop_length) as f32 / sample_rate as f32;
+                let end = if i + 1 < frame_indices.len() {
+                    (frame_indices[i + 1] * encoder_stride * hop_length) as f32 / sample_rate as f32
+                } else {
+                    start + 0.01
+                };
+
+                // Handle SentencePiece format (▁ prefix for word start)
+                let display_text = token_text.replace('▁', " ");
+
+                // Skip special tokens
+                if !(token_text.starts_with('<') && token_text.ends_with('>') && token_text != "<unk>") {
+                    full_text.push_str(&display_text);
+
+                    result_tokens.push(crate::decoder::TimedToken {
+                        text: display_text,
+                        start,
+                        end,
+                    });
+                }
+            }
+        }
+
+        Ok(TranscriptionResult {
+            text: full_text.trim().to_string(),
+            tokens: result_tokens,
+        })
+    }
+}
diff --git a/vendor/parakeet-rs/src/error.rs b/vendor/parakeet-rs/src/error.rs
new file mode 100644
index 0000000..690e0e5
--- /dev/null
+++ b/vendor/parakeet-rs/src/error.rs
@@ -0,0 +1,52 @@
+use std::fmt;
+
+pub type Result<T> = std::result::Result<T, Error>;
+
+#[derive(Debug)]
+pub enum Error {
+    Io(std::io::Error),
+    Ort(ort::Error),
+    Audio(String),
+    Model(String),
+    Tokenizer(String),
+    Config(String),
+}
+
+impl fmt::Display for Error {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            Error::Io(e) => write!(f, "IO error: {e}"),
+            Error::Ort(e) => write!(f, "ONNX Runtime error: {e}"),
+            Error::Audio(msg) => write!(f, "Audio processing error: {msg}"),
+            Error::Model(msg) => write!(f, "Model error: {msg}"),
+            Error::Tokenizer(msg) => write!(f, "Tokenizer error: {msg}"),
+            Error::Config(msg) => write!(f, "Config error: {msg}"),
+        }
+    }
+}
+
+impl std::error::Error for Error {}
+
+impl From<std::io::Error> for Error {
+    fn from(e: std::io::Error) -> Self {
+        Error::Io(e)
+    }
+}
+
+impl From<ort::Error> for Error {
+    fn from(e: ort::Error) -> Self {
+        Error::Ort(e)
+    }
+}
+
+impl From<serde_json::Error> for Error {
+    fn from(e: serde_json::Error) -> Self {
+        Error::Config(e.to_string())
+    }
+}
+
+impl From<hound::Error> for Error {
+    fn from(e: hound::Error) -> Self {
+        Error::Audio(e.to_string())
+    }
+}
diff --git a/vendor/parakeet-rs/src/execution.rs b/vendor/parakeet-rs/src/execution.rs
new file mode 100644
index 0000000..e29aa1d
--- /dev/null
+++ b/vendor/parakeet-rs/src/execution.rs
@@ -0,0 +1,141 @@
+use crate::error::Result;
+use ort::session::builder::SessionBuilder;
+
+// Hardware acceleration options. CPU is default and most reliable.
+// GPU providers (CUDA, TensorRT, ROCm) offer 5-10x speedup but require specific hardware.
+// All GPU providers automatically fall back to CPU if they fail.
+//
+// Note: CoreML currently fails with this model due to unsupported operations.
+// WebGPU is experimental and may produce incorrect results.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum ExecutionProvider {
+    #[default]
+    Cpu,
+    #[cfg(feature = "cuda")]
+    Cuda,
+    #[cfg(feature = "tensorrt")]
+    TensorRT,
+    #[cfg(feature = "coreml")]
+    CoreML,
+    #[cfg(feature = "directml")]
+    DirectML,
+    #[cfg(feature = "rocm")]
+    ROCm,
+    #[cfg(feature = "openvino")]
+    OpenVINO,
+    #[cfg(feature = "webgpu")]
+    WebGPU,
+}
+
+#[derive(Debug, Clone)]
+pub struct ModelConfig {
+    pub execution_provider: ExecutionProvider,
+    pub intra_threads: usize,
+    pub inter_threads: usize,
+}
+
+impl Default for ModelConfig {
+    fn default() -> Self {
+        Self {
+            execution_provider: ExecutionProvider::default(),
+            intra_threads: 4,
+            inter_threads: 1,
+        }
+    }
+}
+
+impl ModelConfig {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn with_execution_provider(mut self, provider: ExecutionProvider) -> Self {
+        self.execution_provider = provider;
+        self
+    }
+
+    pub fn with_intra_threads(mut self, threads: usize) -> Self {
+        self.intra_threads = threads;
+        self
+    }
+
+    pub fn with_inter_threads(mut self, threads: usize) -> Self {
+        self.inter_threads = threads;
+        self
+    }
+
+    pub(crate) fn apply_to_session_builder(
+        &self,
+        builder: SessionBuilder,
+    ) -> Result<SessionBuilder> {
+        use ort::session::builder::GraphOptimizationLevel;
+        #[cfg(any(
+            feature = "cuda",
+            feature = "tensorrt",
+            feature = "coreml",
+            feature = "directml",
+            feature = "rocm",
+            feature = "openvino",
+            feature = "webgpu"
+        ))]
+        use ort::execution_providers::CPUExecutionProvider;
+
+        let mut builder = builder
+            .with_optimization_level(GraphOptimizationLevel::Level3)?
+            .with_intra_threads(self.intra_threads)?
+            .with_inter_threads(self.inter_threads)?;
+
+        builder = match self.execution_provider {
+            ExecutionProvider::Cpu => builder,
+
+            #[cfg(feature = "cuda")]
+            ExecutionProvider::Cuda => builder.with_execution_providers([
+                ort::execution_providers::CUDAExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+
+            #[cfg(feature = "tensorrt")]
+            ExecutionProvider::TensorRT => builder.with_execution_providers([
+                ort::execution_providers::TensorRTExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+
+            #[cfg(feature = "coreml")]
+            ExecutionProvider::CoreML => {
+                use ort::execution_providers::coreml::{CoreMLComputeUnits, CoreMLExecutionProvider};
+                builder.with_execution_providers([
+                    CoreMLExecutionProvider::default()
+                        .with_compute_units(CoreMLComputeUnits::CPUAndGPU)
+                        .build(),
+                    CPUExecutionProvider::default().build().error_on_failure(),
+                ])?
+            }
+
+            #[cfg(feature = "directml")]
+            ExecutionProvider::DirectML => builder.with_execution_providers([
+                ort::execution_providers::DirectMLExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+
+            #[cfg(feature = "rocm")]
+            ExecutionProvider::ROCm => builder.with_execution_providers([
+                ort::execution_providers::ROCMExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+
+            #[cfg(feature = "openvino")]
+            ExecutionProvider::OpenVINO => builder.with_execution_providers([
+                ort::execution_providers::OpenVINOExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+
+            #[cfg(feature = "webgpu")]
+            ExecutionProvider::WebGPU => builder.with_execution_providers([
+                ort::execution_providers::WebGPUExecutionProvider::default().build(),
+                CPUExecutionProvider::default().build().error_on_failure(),
+            ])?,
+        };
+
+        Ok(builder)
+    }
+}
diff --git a/vendor/parakeet-rs/src/lib.rs b/vendor/parakeet-rs/src/lib.rs
new file mode 100644
index 0000000..0aaefd1
--- /dev/null
+++ b/vendor/parakeet-rs/src/lib.rs
@@ -0,0 +1,74 @@
+//! # parakeet-rs
+//!
+//! Rust bindings for NVIDIA's Parakeet speech recognition model using ONNX Runtime.
+//!
+//! Parakeet is a state-of-the-art automatic speech recognition (ASR) model developed by NVIDIA,
+//! based on the FastConformer-TDT architecture with 600 million parameters.
+//!
+//! ## Features
+//!
+//! - Easy-to-use API for speech-to-text transcription
+//! - Support for ONNX format models
+//! - 16kHz mono audio input
+//! - Punctuation and capitalization included in output
+//! - Fast inference using ONNX Runtime
+//!
+//! ## Quick Start
+//!
+//! ```ignore
+//! use parakeet_rs::Parakeet;
+//!
+//! // Load the model
+//! let parakeet = Parakeet::from_pretrained(".")?;
+//!
+//! // Transcribe audio file
+//! let text = parakeet.transcribe_file("audio.wav")?;
+//! println!("Transcription: {}", text);
+//! ```
+//!
+//! ## Model Requirements
+//!
+//! Your model directory should contain:
+//! - `model.onnx` - The ONNX model file
+//! - `model.onnx_data` - External model weights
+//! - `config.json` - Model configuration
+//! - `preprocessor_config.json` - Audio preprocessing configuration
+//! - `tokenizer.json` - Tokenizer vocabulary
+//! - `tokenizer_config.json` - Tokenizer configuration
+//!
+//! ## Audio Requirements
+//!
+//! - Format: WAV
+//! - Sample Rate: 16kHz
+//! - Channels: Mono (stereo will be converted automatically)
+//! - Bit Depth: 16-bit PCM or 32-bit float
+
+mod audio;
+mod config;
+mod decoder;
+mod decoder_tdt;
+mod error;
+mod execution;
+mod model;
+mod model_tdt;
+mod parakeet;
+mod parakeet_tdt;
+mod timestamps;
+mod vocab;
+mod model_eou;
+mod parakeet_eou;
+#[cfg(feature = "sortformer")]
+pub mod sortformer;
+
+pub use error::{Error, Result};
+pub use execution::{ExecutionProvider, ModelConfig as ExecutionConfig};
+pub use parakeet::Parakeet;
+pub use parakeet_tdt::ParakeetTDT;
+pub use timestamps::TimestampMode;
+
+pub use config::{ModelConfig as ModelConfigJson, PreprocessorConfig};
+
+pub use decoder::{ParakeetDecoder, TimedToken, TranscriptionResult};
+pub use model::ParakeetModel;
+pub use model_eou::ParakeetEOUModel;
+pub use parakeet_eou::ParakeetEOU;
\ No newline at end of file
diff --git a/vendor/parakeet-rs/src/model.rs b/vendor/parakeet-rs/src/model.rs
new file mode 100644
index 0000000..b3cd131
--- /dev/null
+++ b/vendor/parakeet-rs/src/model.rs
@@ -0,0 +1,93 @@
+use crate::config::ModelConfig;
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use ndarray::Array2;
+use ort::session::Session;
+use std::path::Path;
+
+pub struct ParakeetModel {
+    session: Session,
+    config: ModelConfig,
+}
+
+impl ParakeetModel {
+    pub fn from_pretrained<P: AsRef<Path>>(model_path: P) -> Result<Self> {
+        Self::from_pretrained_with_config(model_path, ExecutionConfig::default())
+    }
+
+    pub fn from_pretrained_with_config<P: AsRef<Path>>(
+        model_path: P,
+        exec_config: ExecutionConfig,
+    ) -> Result<Self> {
+        let model_path = model_path.as_ref();
+
+        // Use default config (hardcoded constants for Parakeet-CTC-0.6b: please see: json files https://huggingface.co/onnx-community/parakeet-ctc-0.6b-ONNX/tree/main)
+        let config = ModelConfig::default();
+
+        let builder = Session::builder()?;
+        let builder = exec_config.apply_to_session_builder(builder)?;
+        let session = builder.commit_from_file(model_path)?;
+
+        Ok(Self { session, config })
+    }
+    pub fn forward(&mut self, features: Array2<f32>) -> Result<Array2<f32>> {
+        let batch_size = 1;
+        let time_steps = features.shape()[0];
+        let feature_size = features.shape()[1];
+
+        let input = features
+            .to_shape((batch_size, time_steps, feature_size))
+            .map_err(|e| Error::Model(format!("Failed to reshape input: {e}")))?
+            .to_owned();
+
+        use ndarray::Array2;
+        let attention_mask = Array2::<i64>::ones((batch_size, time_steps));
+
+        let input_value = ort::value::Value::from_array(input)?;
+        let attention_mask_value = ort::value::Value::from_array(attention_mask)?;
+
+        let outputs = self.session.run(ort::inputs!(
+            "input_features" => input_value,
+            "attention_mask" => attention_mask_value
+        ))?;
+
+        let logits_value = &outputs["logits"];
+        let (shape, data) = logits_value
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract logits: {e}")))?;
+
+        let shape_dims = shape.as_ref();
+        if shape_dims.len() != 3 {
+            return Err(Error::Model(format!(
+                "Expected 3D logits, got shape: {shape_dims:?}"
+            )));
+        }
+
+        let batch_size = shape_dims[0] as usize;
+        let time_steps_out = shape_dims[1] as usize;
+        let vocab_size = shape_dims[2] as usize;
+
+        if batch_size != 1 {
+            return Err(Error::Model(format!(
+                "Expected batch size 1, got {batch_size}"
+            )));
+        }
+
+        let logits_2d = Array2::from_shape_vec((time_steps_out, vocab_size), data.to_vec())
+            .map_err(|e| Error::Model(format!("Failed to create array: {e}")))?;
+
+        Ok(logits_2d)
+    }
+
+    pub fn config(&self) -> &ModelConfig {
+        &self.config
+    }
+
+    pub fn vocab_size(&self) -> usize {
+        self.config.vocab_size
+    }
+
+    pub fn pad_token_id(&self) -> usize {
+        self.config.pad_token_id
+    }
+}
diff --git a/vendor/parakeet-rs/src/model_eou.rs b/vendor/parakeet-rs/src/model_eou.rs
new file mode 100644
index 0000000..5b56e6d
--- /dev/null
+++ b/vendor/parakeet-rs/src/model_eou.rs
@@ -0,0 +1,183 @@
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use ndarray::{Array1, Array2, Array3, Array4};
+use ort::session::Session;
+use std::path::Path;
+
+/// Encoder cache state for streaming inference
+/// The cache maintains temporal context across chunks
+pub struct EncoderCache {
+    /// channel cache: [1, 1, 70, 512] - batch=1, 70 frame lookback
+    pub cache_last_channel: Array4<f32>,
+    /// time cache: [1, 1, 512, 8] - batch=1, fixed 8 time steps
+    pub cache_last_time: Array4<f32>,
+    /// cache length: [1] with value 0 initially
+    pub cache_last_channel_len: Array1<i64>,
+}
+
+impl EncoderCache {
+    /// 17 layers, batch=1, 70 frame lookback, 512 features
+    pub fn new() -> Self {
+        Self {
+            cache_last_channel: Array4::zeros((17, 1, 70, 512)),
+            cache_last_time: Array4::zeros((17, 1, 512, 8)),
+            cache_last_channel_len: Array1::from_vec(vec![0i64]),
+        }
+    }
+}
+
+pub struct ParakeetEOUModel {
+    encoder: Session,
+    decoder_joint: Session,
+}
+
+impl ParakeetEOUModel {
+    pub fn from_pretrained<P: AsRef<Path>>(
+        model_dir: P,
+        exec_config: ExecutionConfig,
+    ) -> Result<Self> {
+        let model_dir = model_dir.as_ref();
+
+        let encoder_path = model_dir.join("encoder.onnx");
+        let decoder_path = model_dir.join("decoder_joint.onnx");
+
+        if !encoder_path.exists() || !decoder_path.exists() {
+             return Err(Error::Config(format!(
+                "Missing ONNX files in {}. Expected encoder.onnx and decoder_joint.onnx",
+                model_dir.display()
+            )));
+        }
+
+        // Load encoder
+        let builder = Session::builder()?;
+        let builder = exec_config.apply_to_session_builder(builder)?;
+        let encoder = builder.commit_from_file(&encoder_path)?;
+
+        // Load decoder
+        let builder = Session::builder()?;
+        let builder = exec_config.apply_to_session_builder(builder)?;
+        let decoder_joint = builder.commit_from_file(&decoder_path)?;
+
+        Ok(Self {
+            encoder,
+            decoder_joint,
+        })
+    }
+
+    /// Run the stateful encoder with cache
+    /// Input: features [1, 128, T], cache state
+    /// Output: (encoded [1, 512, T], new_cache)
+    pub fn run_encoder(
+        &mut self,
+        features: &Array3<f32>,
+        length: i64,
+        cache: &EncoderCache
+    ) -> Result<(Array3<f32>, EncoderCache)> {
+        let length_arr = Array1::from_vec(vec![length]);
+
+        let outputs = self.encoder.run(ort::inputs![
+            "audio_signal" => ort::value::Value::from_array(features.clone())?,
+            "length" => ort::value::Value::from_array(length_arr)?,
+            "cache_last_channel" => ort::value::Value::from_array(cache.cache_last_channel.clone())?,
+            "cache_last_time" => ort::value::Value::from_array(cache.cache_last_time.clone())?,
+            "cache_last_channel_len" => ort::value::Value::from_array(cache.cache_last_channel_len.clone())?
+        ])?;
+
+        // Extract encoder output [1, 512, T]
+        let (shape, data) = outputs["outputs"]
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract encoder output: {e}")))?;
+
+        let shape_dims = shape.as_ref();
+        let b = shape_dims[0] as usize;
+        let d = shape_dims[1] as usize;
+        let t = shape_dims[2] as usize;
+
+        let encoder_out = Array3::from_shape_vec((b, d, t), data.to_vec())
+            .map_err(|e| Error::Model(format!("Failed to reshape encoder output: {e}")))?;
+
+        // Extract new cache states
+        let (ch_shape, ch_data) = outputs["new_cache_last_channel"]
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract cache_last_channel: {e}")))?;
+
+        let (tm_shape, tm_data) = outputs["new_cache_last_time"]
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract cache_last_time: {e}")))?;
+
+        let (len_shape, len_data) = outputs["new_cache_last_channel_len"]
+            .try_extract_tensor::<i64>()
+            .map_err(|e| Error::Model(format!("Failed to extract cache_len: {e}")))?;
+
+        // Build new cache with extracted shapes
+        let new_cache = EncoderCache {
+            cache_last_channel: Array4::from_shape_vec(
+                (ch_shape[0] as usize, ch_shape[1] as usize, ch_shape[2] as usize, ch_shape[3] as usize),
+                ch_data.to_vec()
+            ).map_err(|e| Error::Model(format!("Failed to reshape cache_last_channel: {e}")))?,
+
+            cache_last_time: Array4::from_shape_vec(
+                (tm_shape[0] as usize, tm_shape[1] as usize, tm_shape[2] as usize, tm_shape[3] as usize),
+                tm_data.to_vec()
+            ).map_err(|e| Error::Model(format!("Failed to reshape cache_last_time: {e}")))?,
+
+            cache_last_channel_len: Array1::from_shape_vec(
+                len_shape[0] as usize,
+                len_data.to_vec()
+            ).map_err(|e| Error::Model(format!("Failed to reshape cache_len: {e}")))?,
+        };
+
+        Ok((encoder_out, new_cache))
+    }
+
+    /// Run the stateful decoder
+    /// Returns: (logits [1, 1, 1, vocab], new_state_h, new_state_c)
+    pub fn run_decoder(
+        &mut self,
+        encoder_frame: &Array3<f32>, // [1, 512, 1]
+        last_token: &Array2<i32>,    // [1, 1]
+        state_h: &Array3<f32>,       // [1, 1, 640]
+        state_c: &Array3<f32>,       // [1, 1, 640]
+    ) -> Result<(Array3<f32>, Array3<f32>, Array3<f32>)> {
+
+        // Target length is always 1 for single step
+        let target_len = Array1::from_vec(vec![1i32]);
+
+        let outputs = self.decoder_joint.run(ort::inputs![
+            "encoder_outputs" => ort::value::Value::from_array(encoder_frame.clone())?,
+            "targets" => ort::value::Value::from_array(last_token.clone())?,
+            "target_length" => ort::value::Value::from_array(target_len)?,
+            "input_states_1" => ort::value::Value::from_array(state_h.clone())?,
+            "input_states_2" => ort::value::Value::from_array(state_c.clone())?
+        ])?;
+
+        // 1. Extract Logits
+        let (l_shape, l_data) = outputs["outputs"]
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract logits: {e}")))?;
+        
+        // 2. Extract States (output_states_1, output_states_2)
+        let (_h_shape, h_data) = outputs["output_states_1"]
+             .try_extract_tensor::<f32>()
+             .map_err(|e| Error::Model(format!("Failed to extract state h: {e}")))?;
+
+        let (_c_shape, c_data) = outputs["output_states_2"]
+             .try_extract_tensor::<f32>()
+             .map_err(|e| Error::Model(format!("Failed to extract state c: {e}")))?;
+
+        // Reconstruct Arrays
+        // Logits: I simplify to [1, 1, vocab]
+        let vocab_size = l_shape[3] as usize;
+        let logits = Array3::from_shape_vec((1, 1, vocab_size), l_data.to_vec())
+            .map_err(|e| Error::Model(format!("Reshape logits failed: {e}")))?;
+
+        // States: [1, 1, 640]
+        let new_h = Array3::from_shape_vec((1, 1, 640), h_data.to_vec())
+             .map_err(|e| Error::Model(format!("Reshape state h failed: {e}")))?;
+        
+        let new_c = Array3::from_shape_vec((1, 1, 640), c_data.to_vec())
+             .map_err(|e| Error::Model(format!("Reshape state c failed: {e}")))?;
+
+        Ok((logits, new_h, new_c))
+    }
+}
\ No newline at end of file
diff --git a/vendor/parakeet-rs/src/model_tdt.rs b/vendor/parakeet-rs/src/model_tdt.rs
new file mode 100644
index 0000000..e00ebdc
--- /dev/null
+++ b/vendor/parakeet-rs/src/model_tdt.rs
@@ -0,0 +1,263 @@
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use ndarray::{Array1, Array2, Array3};
+use ort::session::Session;
+use std::path::{Path, PathBuf};
+
+/// TDT model configs
+#[derive(Debug, Clone)]
+pub struct TDTModelConfig {
+    pub vocab_size: usize,
+}
+
+impl Default for TDTModelConfig {
+    fn default() -> Self {
+        Self {
+            vocab_size: 8193,
+        }
+    }
+}
+
+pub struct ParakeetTDTModel {
+    encoder: Session,
+    decoder_joint: Session,
+    config: TDTModelConfig,
+}
+
+impl ParakeetTDTModel {
+    /// Load TDT model from directory containing encoder and decoder_joint ONNX files
+    pub fn from_pretrained<P: AsRef<Path>>(
+        model_dir: P,
+        exec_config: ExecutionConfig,
+    ) -> Result<Self> {
+        let model_dir = model_dir.as_ref();
+
+        // Find encoder and decoder_joint files
+        let encoder_path = Self::find_encoder(model_dir)?;
+        let decoder_joint_path = Self::find_decoder_joint(model_dir)?;
+
+        let config = TDTModelConfig::default();
+
+        // Load encoder
+        let builder = Session::builder()?;
+        let builder = exec_config.apply_to_session_builder(builder)?;
+        let encoder = builder.commit_from_file(&encoder_path)?;
+
+        // Load decoder_joint
+        let builder = Session::builder()?;
+        let builder = exec_config.apply_to_session_builder(builder)?;
+        let decoder_joint = builder.commit_from_file(&decoder_joint_path)?;
+
+
+        Ok(Self {
+            encoder,
+            decoder_joint,
+            config,
+        })
+    }
+
+    fn find_encoder(dir: &Path) -> Result<PathBuf> {
+        let candidates = ["encoder-model.onnx", "encoder.onnx"];
+        for candidate in &candidates {
+            let path = dir.join(candidate);
+            if path.exists() {
+                return Ok(path);
+            }
+        }
+        Err(Error::Config(format!(
+            "No encoder model found in {}",
+            dir.display()
+        )))
+    }
+
+    fn find_decoder_joint(dir: &Path) -> Result<PathBuf> {
+        let candidates = [
+            "decoder_joint-model.onnx",
+            "decoder_joint.onnx",
+            "decoder-model.onnx",
+        ];
+        for candidate in &candidates {
+            let path = dir.join(candidate);
+            if path.exists() {
+                return Ok(path);
+            }
+        }
+        Err(Error::Config(format!(
+            "No decoder_joint model found in {}",
+            dir.display()
+        )))
+    }
+
+    /// Run greedy decoding - returns (token_ids, frame_indices, durations)
+    pub fn forward(&mut self, features: Array2<f32>) -> Result<(Vec<usize>, Vec<usize>, Vec<usize>)> {
+        // Run encoder
+        let (encoder_out, encoder_len) = self.run_encoder(&features)?;
+
+        // Run greedy decoding with decoder_joint
+        let (tokens, frame_indices, durations) = self.greedy_decode(&encoder_out, encoder_len)?;
+
+        Ok((tokens, frame_indices, durations))
+    }
+
+    fn run_encoder(&mut self, features: &Array2<f32>) -> Result<(Array3<f32>, i64)> {
+        let batch_size = 1;
+        let time_steps = features.shape()[0];
+        let feature_size = features.shape()[1];
+
+        // TDT encoder expects (batch, features, time) not (batch, time, features)
+        let input = features
+            .t()
+            .to_shape((batch_size, feature_size, time_steps))
+            .map_err(|e| Error::Model(format!("Failed to reshape encoder input: {e}")))?
+            .to_owned();
+
+        let input_length = Array1::from_vec(vec![time_steps as i64]);
+
+        let input_value = ort::value::Value::from_array(input)?;
+        let length_value = ort::value::Value::from_array(input_length)?;
+
+        let outputs = self.encoder.run(ort::inputs!(
+            "audio_signal" => input_value,
+            "length" => length_value
+        ))?;
+
+        let encoder_out = &outputs["outputs"];
+        let encoder_lens = &outputs["encoded_lengths"];
+
+        let (shape, data) = encoder_out
+            .try_extract_tensor::<f32>()
+            .map_err(|e| Error::Model(format!("Failed to extract encoder output: {e}")))?;
+
+        let (_, lens_data) = encoder_lens
+            .try_extract_tensor::<i64>()
+            .map_err(|e| Error::Model(format!("Failed to extract encoder lengths: {e}")))?;
+
+        let shape_dims = shape.as_ref();
+        if shape_dims.len() != 3 {
+            return Err(Error::Model(format!(
+                "Expected 3D encoder output, got shape: {shape_dims:?}"
+            )));
+        }
+
+        let b = shape_dims[0] as usize;
+        let t = shape_dims[1] as usize;
+        let d = shape_dims[2] as usize;
+
+        let encoder_array = Array3::from_shape_vec((b, t, d), data.to_vec())
+            .map_err(|e| Error::Model(format!("Failed to create encoder array: {e}")))?;
+
+        // TDT encoder outputs [batch, encoder_dim, time] directly
+        Ok((encoder_array, lens_data[0]))
+    }
+
+    fn greedy_decode(&mut self, encoder_out: &Array3<f32>, _encoder_len: i64) -> Result<(Vec<usize>, Vec<usize>, Vec<usize>)> {
+        // encoder_out shape: [batch, encoder_dim, time]
+        let encoder_dim = encoder_out.shape()[1];
+        let time_steps = encoder_out.shape()[2];
+        let vocab_size = self.config.vocab_size;
+        let max_tokens_per_step = 10;
+        let blank_id = vocab_size - 1;
+
+        // States: (num_layers=2, batch=1, hidden_dim=640)
+        let mut state_h = Array3::<f32>::zeros((2, 1, 640));
+        let mut state_c = Array3::<f32>::zeros((2, 1, 640));
+
+        let mut tokens = Vec::new();
+        let mut frame_indices = Vec::new();
+        let mut durations = Vec::new();
+
+        let mut t = 0;
+        let mut emitted_tokens = 0;
+        let mut last_emitted_token = blank_id as i32;
+
+        // Frame-by-frame RNN-T/TDT greedy decoding
+        while t < time_steps {
+            // Get single encoder frame: slice [0, :, t] and reshape to [1, encoder_dim, 1]
+            let frame = encoder_out.slice(ndarray::s![0, .., t]).to_owned();
+            let frame_reshaped = frame
+                .to_shape((1, encoder_dim, 1))
+                .map_err(|e| Error::Model(format!("Failed to reshape frame: {e}")))?
+                .to_owned();
+
+            // Current token for prediction network
+            let targets = Array2::from_shape_vec((1, 1), vec![last_emitted_token])
+                .map_err(|e| Error::Model(format!("Failed to create targets: {e}")))?;
+
+            // Run decoder_joint
+            let outputs = self.decoder_joint.run(ort::inputs!(
+                "encoder_outputs" => ort::value::Value::from_array(frame_reshaped)?,
+                "targets" => ort::value::Value::from_array(targets)?,
+                "target_length" => ort::value::Value::from_array(Array1::from_vec(vec![1i32]))?,
+                "input_states_1" => ort::value::Value::from_array(state_h.clone())?,
+                "input_states_2" => ort::value::Value::from_array(state_c.clone())?
+            ))?;
+
+            // Extract logits
+            let (_, logits_data) = outputs["outputs"]
+                .try_extract_tensor::<f32>()
+                .map_err(|e| Error::Model(format!("Failed to extract logits: {e}")))?;
+
+            // TDT outputs vocab_size + 5 durations (8193 + 5 = 8198)
+            let vocab_logits: Vec<f32> = logits_data.iter().take(vocab_size).copied().collect();
+            let duration_logits: Vec<f32> = logits_data.iter().skip(vocab_size).copied().collect();
+
+            let token_id = vocab_logits
+                .iter()
+                .enumerate()
+                .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+                .map(|(idx, _)| idx)
+                .unwrap_or(blank_id);
+
+            let duration_step = if !duration_logits.is_empty() {
+                duration_logits
+                    .iter()
+                    .enumerate()
+                    .max_by(|(_, a), (_, b)| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
+                    .map(|(idx, _)| idx)
+                    .unwrap_or(0)
+            } else {
+                0
+            };
+
+            // Check if blank token
+            if token_id != blank_id {
+                // Update states when we emit a token
+                if let Ok((h_shape, h_data)) = outputs["output_states_1"].try_extract_tensor::<f32>() {
+                    let dims = h_shape.as_ref();
+                    state_h = Array3::from_shape_vec((dims[0] as usize, dims[1] as usize, dims[2] as usize), h_data.to_vec())
+                        .map_err(|e| Error::Model(format!("Failed to update state_h: {e}")))?;
+                }
+                if let Ok((c_shape, c_data)) = outputs["output_states_2"].try_extract_tensor::<f32>() {
+                    let dims = c_shape.as_ref();
+                    state_c = Array3::from_shape_vec((dims[0] as usize, dims[1] as usize, dims[2] as usize), c_data.to_vec())
+                        .map_err(|e| Error::Model(format!("Failed to update state_c: {e}")))?;
+                }
+
+                tokens.push(token_id);
+                frame_indices.push(t);
+                durations.push(duration_step);
+                last_emitted_token = token_id as i32;
+                emitted_tokens += 1;
+
+                // Don't advance yet - try to emit more tokens from the same frame
+            } else {
+                // Blank token - advance frame pointer
+                // Duration prediction applies when we finally move to next frame after emitting tokens
+                if duration_step > 0 && emitted_tokens > 0 {
+                    t += duration_step;
+                } else {
+                    t += 1;
+                }
+                emitted_tokens = 0;
+            }
+
+            // Safety check: if we've emitted too many tokens from the same frame, advance
+            if emitted_tokens >= max_tokens_per_step {
+                t += 1;
+                emitted_tokens = 0;
+            }
+        }
+
+        Ok((tokens, frame_indices, durations))
+    }
+}
diff --git a/vendor/parakeet-rs/src/parakeet.rs b/vendor/parakeet-rs/src/parakeet.rs
new file mode 100644
index 0000000..d2aabdd
--- /dev/null
+++ b/vendor/parakeet-rs/src/parakeet.rs
@@ -0,0 +1,210 @@
+use crate::audio;
+use crate::config::PreprocessorConfig;
+use crate::decoder::{ParakeetDecoder, TranscriptionResult};
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use crate::model::ParakeetModel;
+use crate::timestamps::{process_timestamps, TimestampMode};
+use std::path::{Path, PathBuf};
+
+pub struct Parakeet {
+    model: ParakeetModel,
+    decoder: ParakeetDecoder,
+    preprocessor_config: PreprocessorConfig,
+    model_dir: PathBuf,
+}
+
+impl Parakeet {
+    /// Load Parakeet model from path with optional configuration.
+    ///
+    /// # Arguments
+    /// * `path` - Directory containing model files, or path to specific model file
+    /// * `config` - Optional execution configuration (defaults to CPU if None)
+    ///
+    /// # Examples
+    /// ```no_run
+    /// use parakeet_rs::Parakeet;
+    ///
+    /// // Load from directory with CPU (default)
+    /// let parakeet = Parakeet::from_pretrained(".", None)?;
+    ///
+    /// // Or load from specific model file
+    /// let parakeet = Parakeet::from_pretrained("model_q4.onnx", None)?;
+    /// # Ok::<(), Box<dyn std::error::Error>>(())
+    /// ```
+    ///
+    /// For GPU acceleration, enable the corresponding feature (cuda, tensorrt, webgpu, etc.)
+    /// and pass an `ExecutionConfig` with the desired execution provider.
+    pub fn from_pretrained<P: AsRef<Path>>(
+        path: P,
+        config: Option<ExecutionConfig>,
+    ) -> Result<Self> {
+        let path = path.as_ref();
+
+        // Determine if path is a directory or file
+        let (model_path, tokenizer_path, model_dir) = if path.is_dir() {
+            // Directory mode: auto-detect model file
+            let model_path = Self::find_model_file(path)?;
+            let tokenizer_path = path.join("tokenizer.json");
+            (model_path, tokenizer_path, path.to_path_buf())
+        } else if path.is_file() {
+            // File mode: path points directly to model file
+            let model_dir = path
+                .parent()
+                .ok_or_else(|| Error::Config("Invalid model path".to_string()))?;
+            let tokenizer_path = model_dir.join("tokenizer.json");
+            (path.to_path_buf(), tokenizer_path, model_dir.to_path_buf())
+        } else {
+            return Err(Error::Config(format!(
+                "Path does not exist: {}",
+                path.display()
+            )));
+        };
+
+        // Check tokenizer exists
+        if !tokenizer_path.exists() {
+            return Err(Error::Config(format!(
+                "Required file 'tokenizer.json' not found in {}",
+                model_dir.display()
+            )));
+        }
+
+        let preprocessor_config = PreprocessorConfig::default();
+        let exec_config = config.unwrap_or_default();
+
+        let model = ParakeetModel::from_pretrained_with_config(&model_path, exec_config)?;
+        let decoder = ParakeetDecoder::from_pretrained(&tokenizer_path)?;
+
+        Ok(Self {
+            model,
+            decoder,
+            preprocessor_config,
+            model_dir,
+        })
+    }
+
+    fn find_model_file(dir: &Path) -> Result<PathBuf> {
+        // Priority order: model.onnx > model_fp16.onnx > model_int8.onnx > model_q4.onnx
+        let candidates = [
+            "model.onnx",
+            "model_fp16.onnx",
+            "model_int8.onnx",
+            "model_q4.onnx",
+        ];
+
+        for candidate in &candidates {
+            let path = dir.join(candidate);
+            if path.exists() {
+                return Ok(path);
+            }
+        }
+
+        // If none of the standard names found, search for any .onnx file
+        if let Ok(entries) = std::fs::read_dir(dir) {
+            for entry in entries.flatten() {
+                let path = entry.path();
+                if path.extension().and_then(|s| s.to_str()) == Some("onnx") {
+                    return Ok(path);
+                }
+            }
+        }
+
+        Err(Error::Config(format!(
+            "No model file (*.onnx) found in directory: {}",
+            dir.display()
+        )))
+    }
+
+    /// Transcribe audio samples.
+    ///
+    /// # Arguments
+    ///
+    /// * `audio` - Audio samples as f32 values
+    /// * `sample_rate` - Sample rate in Hz
+    /// * `channels` - Number of audio channels
+    /// * `mode` - Optional timestamp output mode (Tokens, Words, or Sentences)
+    ///
+    /// # Returns
+    ///
+    /// A `TranscriptionResult` containing the transcribed text and timestamps at the requested level.
+    pub fn transcribe_samples(
+        &mut self,
+        audio: Vec<f32>,
+        sample_rate: u32,
+        channels: u16,
+        mode: Option<TimestampMode>,
+    ) -> Result<TranscriptionResult> {
+        let features = audio::extract_features_raw(audio, sample_rate, channels, &self.preprocessor_config)?;
+        let logits = self.model.forward(features)?;
+
+        let mut result = self.decoder.decode_with_timestamps(
+            &logits,
+            self.preprocessor_config.hop_length,
+            self.preprocessor_config.sampling_rate,
+        )?;
+
+        // Process timestamps to requested output mode
+        let mode = mode.unwrap_or(TimestampMode::Tokens);
+        result.tokens = process_timestamps(&result.tokens, mode);
+
+        // Rebuild full text from processed tokens to ensure consistency
+        result.text = result.tokens.iter()
+            .map(|t| t.text.as_str())
+            .collect::<Vec<_>>()
+            .join(" ");
+
+        Ok(result)
+    }
+
+    /// Transcribe an audio file with timestamps
+    ///
+    /// # Arguments
+    ///
+    /// * `audio_path` - A path to the audio file that needs to be transcribed.
+    /// * `mode` - Optional timestamp output mode (Tokens, Words, or Sentences)
+    ///
+    /// # Returns
+    ///
+    /// This function returns a `TranscriptionResult` which includes the transcribed text along with timestamps at the requested level.
+    pub fn transcribe_file<P: AsRef<Path>>(
+        &mut self,
+        audio_path: P,
+        mode: Option<TimestampMode>,
+    ) -> Result<TranscriptionResult> {
+        let audio_path = audio_path.as_ref();
+        let (audio, spec) = audio::load_audio(audio_path)?;
+
+        self.transcribe_samples(audio, spec.sample_rate, spec.channels, mode)
+    }
+
+    /// Transcribes multiple audio files in batch.
+    ///
+    /// # Arguments
+    ///
+    /// * `audio_paths`: A slice of paths to the audio files that need to be transcribed.
+    /// * `mode` - Optional timestamp output mode (Tokens, Words, or Sentences)
+    ///
+    /// # Returns
+    ///
+    /// This function returns a `TranscriptionResult` which includes the transcribed text along with timestamps at the requested level.
+    pub fn transcribe_file_batch<P: AsRef<Path>>(
+        &mut self,
+        audio_paths: &[P],
+        mode: Option<TimestampMode>,
+    ) -> Result<Vec<TranscriptionResult>> {
+        let mut results = Vec::with_capacity(audio_paths.len());
+        for path in audio_paths {
+            let result = self.transcribe_file(path, mode)?;
+            results.push(result);
+        }
+        Ok(results)
+    }
+
+    pub fn model_dir(&self) -> &Path {
+        &self.model_dir
+    }
+
+    pub fn preprocessor_config(&self) -> &PreprocessorConfig {
+        &self.preprocessor_config
+    }
+}
diff --git a/vendor/parakeet-rs/src/parakeet_eou.rs b/vendor/parakeet-rs/src/parakeet_eou.rs
new file mode 100644
index 0000000..25c7d64
--- /dev/null
+++ b/vendor/parakeet-rs/src/parakeet_eou.rs
@@ -0,0 +1,304 @@
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use crate::model_eou::{EncoderCache, ParakeetEOUModel};
+use ndarray::{s, Array2, Array3};
+use rustfft::{num_complex::Complex, FftPlanner};
+use std::collections::VecDeque;
+use std::f32::consts::PI;
+use std::path::Path;
+
+const SAMPLE_RATE: usize = 16000;
+
+const N_FFT: usize = 512;
+const WIN_LENGTH: usize = 400;
+const HOP_LENGTH: usize = 160;
+const N_MELS: usize = 128;
+const PREEMPH: f32 = 0.97;
+const LOG_ZERO_GUARD: f32 = 5.960464478e-8;
+const FMAX: f32 = 8000.0;
+
+/// Parakeet RealTime EOU model for streaming ASR with end-of-utterance detection.
+/// Uses cache-aware streaming with audio buffering for pre-encode context.
+pub struct ParakeetEOU {
+    model: ParakeetEOUModel,
+    tokenizer: tokenizers::Tokenizer,
+    encoder_cache: EncoderCache,
+    state_h: Array3<f32>,
+    state_c: Array3<f32>,
+    last_token: Array2<i32>,
+    blank_id: i32,
+    eou_id: i32,
+    mel_basis: Array2<f32>,
+    window: Vec<f32>,
+    audio_buffer: VecDeque<f32>,
+    buffer_size_samples: usize,
+}
+
+impl ParakeetEOU {
+    /// Load Parakeet EOU model from path
+    ///
+    /// # Arguments
+    /// * `path` - Directory containing encoder.onnx, decoder_joint.onnx, and tokenizer.json
+    /// * `config` - Optional execution configuration (defaults to CPU if None)
+    pub fn from_pretrained<P: AsRef<Path>>(path: P, config: Option<ExecutionConfig>) -> Result<Self> {
+        let path = path.as_ref();
+        let tokenizer_path = path.join("tokenizer.json");
+        let tokenizer = tokenizers::Tokenizer::from_file(&tokenizer_path)
+            .map_err(|e| Error::Config(format!("Failed to load tokenizer: {e}")))?;
+
+        let vocab_size = tokenizer.get_vocab_size(true);
+        let blank_id = (vocab_size - 1) as i32;
+        let blank_id = if blank_id < 1000 { 1026 } else { blank_id };
+        let eou_id = tokenizer.token_to_id("<EOU>").map(|id| id as i32).unwrap_or(1024);
+
+        let exec_config = config.unwrap_or_default();
+        let model = ParakeetEOUModel::from_pretrained(path, exec_config)?;
+
+        // Buffer size: 4 seconds of audio
+        // Provides long history for feature extraction context
+        // Note that, I pick those "magic numbers" by looking NeMo's ring buffer approach.
+        let buffer_size_samples = SAMPLE_RATE * 4;  // 4 seconds = 64000 samples
+
+        Ok(Self {
+            model,
+            tokenizer,
+            encoder_cache: EncoderCache::new(),
+            state_h: Array3::zeros((1, 1, 640)),
+            state_c: Array3::zeros((1, 1, 640)),
+            last_token: Array2::from_elem((1, 1), blank_id),
+            blank_id,
+            eou_id,
+            mel_basis: Self::create_mel_filterbank(),
+            window: Self::create_window(),
+            audio_buffer: VecDeque::with_capacity(buffer_size_samples),
+            buffer_size_samples,
+        })
+    }
+
+    /// Transcribe a chunk of audio samples.
+    ///
+    /// # Arguments
+    /// * `chunk` - Audio chunk (typically 160ms / 2560 samples at 16kHz)
+    /// * `reset_on_eou` - If true, reset decoder state when end-of-utterance is detected
+    ///
+    /// # Streaming Behavior
+    /// Cache-aware streaming
+    /// - Maintains 4-second ring buffer for feature extraction context
+    /// - Extracts features from full buffer
+    /// - Slices last (pre_encode_cache + new_frames) for encoder input
+    /// - pre_encode_cache=9 frames, new_frames=~16, total=~25 frames to encoder
+    pub fn transcribe(&mut self, chunk: &[f32], reset_on_eou: bool) -> Result<String> {
+        // Add new chunk to rolling buffer
+        self.audio_buffer.extend(chunk.iter().copied());
+
+        // Trim buffer to keep only the most recent samples
+        while self.audio_buffer.len() > self.buffer_size_samples {
+            self.audio_buffer.pop_front();
+        }
+
+        // Wait until buffer has minimum samples (at least 1 second for stable features)
+        const MIN_BUFFER_SAMPLES: usize = SAMPLE_RATE;  // 1 second
+        if self.audio_buffer.len() < MIN_BUFFER_SAMPLES {
+            return Ok(String::new());
+        }
+
+        // Extract features from FULL buffer (provides context for feature extraction)
+        let buffer_slice: Vec<f32> = self.audio_buffer.iter().copied().collect();
+        let full_features = self.extract_mel_features(&buffer_slice);
+        let total_frames = full_features.shape()[2];
+
+        // Slice to take only (pre_encode_cache + new_frames) for encoder
+        // pre_encode_cache = 9 frames, new_frames = ~16 for 160ms chunk
+        const PRE_ENCODE_CACHE: usize = 9;
+        const FRAMES_PER_CHUNK: usize = 16;
+        const SLICE_LEN: usize = PRE_ENCODE_CACHE + FRAMES_PER_CHUNK;
+
+        let start_frame = if total_frames > SLICE_LEN {
+            total_frames - SLICE_LEN
+        } else {
+            0
+        };
+
+        let features = full_features.slice(s![.., .., start_frame..]).to_owned();
+        let time_steps = features.shape()[2];
+
+        // Encode with cache - encoder sees full buffer context
+        let (encoder_out, new_cache) = self.model.run_encoder(&features, time_steps as i64, &self.encoder_cache)?;
+        self.encoder_cache = new_cache;
+
+        let total_frames = encoder_out.shape()[2];
+        if total_frames == 0 {
+            return Ok(String::new());
+        }
+
+        // Process all output frames (typically 1 frame per chunk)
+        let new_frames = encoder_out;
+
+        let mut text_output = String::new();
+
+        for t in 0..new_frames.shape()[2] {
+            let current_frame = new_frames.slice(s![.., .., t..t + 1]).to_owned();
+            let mut syms_added = 0;
+
+            while syms_added < 5 {
+                let (logits, new_h, new_c) = self.model.run_decoder(
+                    &current_frame,
+                    &self.last_token,
+                    &self.state_h,
+                    &self.state_c,
+                )?;
+
+                let vocab = logits.slice(s![0, 0, ..]);
+
+                let mut max_idx = 0;
+                let mut max_val = f32::NEG_INFINITY;
+                for (i, &val) in vocab.iter().enumerate() {
+                    if val.is_finite() && val > max_val {
+                        max_val = val;
+                        max_idx = i as i32;
+                    }
+                }
+
+                if max_idx == self.blank_id || max_idx == 0 {
+                    break;
+                }
+
+                if max_idx == self.eou_id {
+                    if reset_on_eou {
+                        self.reset_states();
+                        return Ok(text_output + " [EOU]");
+                    }
+                    break;
+                }
+
+                if max_idx as usize >= self.tokenizer.get_vocab_size(true) {
+                    break;
+                }
+
+                self.state_h = new_h;
+                self.state_c = new_c;
+                self.last_token.fill(max_idx);
+
+                if let Some(token) = self.tokenizer.id_to_token(max_idx as u32) {
+                    let clean = token.replace('▁', " ");
+                    text_output.push_str(&clean);
+                }
+                syms_added += 1;
+            }
+        }
+        Ok(text_output)
+    }
+
+    fn reset_states(&mut self) {
+        // Soft reset: Only reset decoder states
+        // at this state, we need to keep encoder cache and audio buffer flowing for continuous context
+        // self.encoder_cache = EncoderCache::new();  // DON'T reset!!!
+        self.state_h.fill(0.0);
+        self.state_c.fill(0.0);
+        self.last_token.fill(self.blank_id);
+        // self.audio_buffer.clear();  // DON'T clear!! 
+    }
+
+    fn extract_mel_features(&self, audio: &[f32]) -> Array3<f32> {
+        let audio_pre = Self::apply_preemphasis(audio);
+        let spec = self.stft(&audio_pre);
+        let mel = self.mel_basis.dot(&spec);
+        let mel_log = mel.mapv(|x| (x.max(0.0) + LOG_ZERO_GUARD).ln());
+        mel_log.insert_axis(ndarray::Axis(0))
+    }
+
+    fn apply_preemphasis(audio: &[f32]) -> Vec<f32> {
+        let mut result = Vec::with_capacity(audio.len());
+        if audio.is_empty() {
+            return result;
+        }
+
+        let safe_x = |x: f32| if x.is_finite() { x } else { 0.0 };
+
+        result.push(safe_x(audio[0]));
+        for i in 1..audio.len() {
+            result.push(safe_x(audio[i]) - PREEMPH * safe_x(audio[i - 1]));
+        }
+        result
+    }
+
+    fn stft(&self, audio: &[f32]) -> Array2<f32> {
+        let mut planner = FftPlanner::<f32>::new();
+        let fft = planner.plan_fft_forward(N_FFT);
+
+        let pad_amount = N_FFT / 2;
+        let mut padded_audio = vec![0.0; pad_amount];
+        padded_audio.extend_from_slice(audio);
+        padded_audio.extend(std::iter::repeat(0.0).take(pad_amount));
+
+        let num_frames = 1 + (padded_audio.len().saturating_sub(WIN_LENGTH)) / HOP_LENGTH;
+        let freq_bins = N_FFT / 2 + 1;
+        let mut spec = Array2::zeros((freq_bins, num_frames));
+
+        for frame_idx in 0..num_frames {
+            let start = frame_idx * HOP_LENGTH;
+            if start + WIN_LENGTH > padded_audio.len() {
+                break;
+            }
+
+            let mut buffer: Vec<Complex<f32>> = vec![Complex::new(0.0, 0.0); N_FFT];
+            for i in 0..WIN_LENGTH {
+                buffer[i] = Complex::new(padded_audio[start + i] * self.window[i], 0.0);
+            }
+            fft.process(&mut buffer);
+            for (i, val) in buffer.iter().take(freq_bins).enumerate() {
+                let mag_sq = val.norm_sqr();
+                spec[[i, frame_idx]] = if mag_sq.is_finite() { mag_sq } else { 0.0 };
+            }
+        }
+        spec
+    }
+
+    fn create_window() -> Vec<f32> {
+        (0..WIN_LENGTH)
+            .map(|i| 0.5 - 0.5 * ((2.0 * PI * i as f32) / ((WIN_LENGTH - 1) as f32)).cos())
+            .collect()
+    }
+
+    fn create_mel_filterbank() -> Array2<f32> {
+        let num_freqs = N_FFT / 2 + 1;
+
+        let hz_to_mel = |hz: f32| 2595.0 * (1.0 + hz / 700.0).log10();
+        let mel_to_hz = |mel: f32| 700.0 * (10.0_f32.powf(mel / 2595.0) - 1.0);
+
+        let mel_min = hz_to_mel(0.0);
+        let mel_max = hz_to_mel(FMAX);
+
+        let mel_points: Vec<f32> = (0..=N_MELS + 1)
+            .map(|i| mel_to_hz(mel_min + (mel_max - mel_min) * i as f32 / (N_MELS + 1) as f32))
+            .collect();
+
+        let fft_freqs: Vec<f32> = (0..num_freqs)
+            .map(|i| (SAMPLE_RATE as f32 / N_FFT as f32) * i as f32)
+            .collect();
+
+        let mut weights = Array2::zeros((N_MELS, num_freqs));
+
+        for i in 0..N_MELS {
+            let left = mel_points[i];
+            let center = mel_points[i + 1];
+            let right = mel_points[i + 2];
+            for (j, &freq) in fft_freqs.iter().enumerate() {
+                if freq >= left && freq <= center {
+                    weights[[i, j]] = (freq - left) / (center - left);
+                } else if freq > center && freq <= right {
+                    weights[[i, j]] = (right - freq) / (right - center);
+                }
+            }
+        }
+
+        for i in 0..N_MELS {
+            let enorm = 2.0 / (mel_points[i + 2] - mel_points[i]);
+            for j in 0..num_freqs {
+                weights[[i, j]] *= enorm;
+            }
+        }
+
+        weights
+    }
+}
diff --git a/vendor/parakeet-rs/src/parakeet_tdt.rs b/vendor/parakeet-rs/src/parakeet_tdt.rs
new file mode 100644
index 0000000..719ae75
--- /dev/null
+++ b/vendor/parakeet-rs/src/parakeet_tdt.rs
@@ -0,0 +1,167 @@
+use crate::audio;
+use crate::config::PreprocessorConfig;
+use crate::decoder::TranscriptionResult;
+use crate::decoder_tdt::ParakeetTDTDecoder;
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig as ExecutionConfig;
+use crate::model_tdt::ParakeetTDTModel;
+use crate::timestamps::{process_timestamps, TimestampMode};
+use crate::vocab::Vocabulary;
+use std::path::{Path, PathBuf};
+
+/// Parakeet TDT model for multilingual ASR
+pub struct ParakeetTDT {
+    model: ParakeetTDTModel,
+    decoder: ParakeetTDTDecoder,
+    preprocessor_config: PreprocessorConfig,
+    model_dir: PathBuf,
+}
+
+impl ParakeetTDT {
+    /// Load Parakeet TDT model from path with optional configuration.
+    ///
+    /// # Arguments
+    /// * `path` - Directory containing encoder-model.onnx, decoder_joint-model.onnx, and vocab.txt
+    /// * `config` - Optional execution configuration (defaults to CPU if None)
+    pub fn from_pretrained<P: AsRef<Path>>(
+        path: P,
+        config: Option<ExecutionConfig>,
+    ) -> Result<Self> {
+        let path = path.as_ref();
+
+        if !path.is_dir() {
+            return Err(Error::Config(format!(
+                "TDT model path must be a directory: {}",
+                path.display()
+            )));
+        }
+
+        let vocab_path = path.join("vocab.txt");
+        if !vocab_path.exists() {
+            return Err(Error::Config(format!(
+                "vocab.txt not found in {}",
+                path.display()
+            )));
+        }
+
+        // TDT-specific preprocessor config (128 features instead of 80)
+        let preprocessor_config = PreprocessorConfig {
+            feature_extractor_type: "ParakeetFeatureExtractor".to_string(),
+            feature_size: 128,
+            hop_length: 160,
+            n_fft: 512,
+            padding_side: "right".to_string(),
+            padding_value: 0.0,
+            preemphasis: 0.97,
+            processor_class: "ParakeetProcessor".to_string(),
+            return_attention_mask: true,
+            sampling_rate: 16000,
+            win_length: 400,
+        };
+
+        let exec_config = config.unwrap_or_default();
+
+        let model = ParakeetTDTModel::from_pretrained(path, exec_config)?;
+        let vocab = Vocabulary::from_file(&vocab_path)?;
+        let decoder = ParakeetTDTDecoder::from_vocab(vocab);
+
+        Ok(Self {
+            model,
+            decoder,
+            preprocessor_config,
+            model_dir: path.to_path_buf(),
+        })
+    }
+
+    /// Transcribe audio samples.
+    ///
+    /// # Arguments
+    ///
+    /// * `audio` - Audio samples as f32 values
+    /// * `sample_rate` - Sample rate in Hz
+    /// * `channels` - Number of audio channels
+    /// * `mode` - Optional timestamp mode (Token, Word, or Segment)
+    ///
+    /// # Returns
+    ///
+    /// A `TranscriptionResult` containing the transcribed text and timestamps at the requested mode.
+    pub fn transcribe_samples(
+        &mut self,
+        audio: Vec<f32>,
+        sample_rate: u32,
+        channels: u16,
+        mode: Option<TimestampMode>,
+    ) -> Result<TranscriptionResult> {
+        let features = audio::extract_features_raw(audio, sample_rate, channels, &self.preprocessor_config)?;
+        let (tokens, frame_indices, durations) = self.model.forward(features)?;
+
+        let mut result = self.decoder.decode_with_timestamps(
+            &tokens,
+            &frame_indices,
+            &durations,
+            self.preprocessor_config.hop_length,
+            self.preprocessor_config.sampling_rate,
+        )?;
+
+        // Apply timestamp mode conversion
+        let mode = mode.unwrap_or(TimestampMode::Tokens);
+        result.tokens = process_timestamps(&result.tokens, mode);
+
+        // Rebuild full text from processed tokens
+        result.text = result.tokens.iter()
+            .map(|t| t.text.as_str())
+            .collect::<Vec<_>>()
+            .join(" ");
+
+        Ok(result)
+    }
+
+    /// Transcribe an audio file with timestamps
+    ///
+    /// # Arguments
+    ///
+    /// * `audio_path` - A path to the audio file that needs to be transcribed.
+    /// * `mode` - Optional timestamp mode (Token, Word, or Segment)
+    ///
+    /// # Returns
+    ///
+    /// This function returns a `TranscriptionResult` which includes the transcribed text along with timestamps at the requested mode.
+    pub fn transcribe_file<P: AsRef<Path>>(
+        &mut self,
+        audio_path: P,
+        mode: Option<TimestampMode>,
+    ) -> Result<TranscriptionResult> {
+        let audio_path = audio_path.as_ref();
+        let (audio, spec) = audio::load_audio(audio_path)?;
+
+        self.transcribe_samples(audio, spec.sample_rate, spec.channels, mode)
+    }
+
+    /// Transcribes multiple audio files in batch.
+    ///
+    /// # Arguments
+    ///
+    /// * `audio_paths`: A slice of paths to the audio files that need to be transcribed.
+    /// * `mode` - Optional timestamp mode (Token, Word, or Segment)
+    ///
+    /// # Returns
+    ///
+    /// This function returns a `TranscriptionResult` which includes the transcribed text along with timestamps at the requested mode.
+    pub fn transcribe_file_batch<P: AsRef<Path>>(
+        &mut self,
+        audio_paths: &[P],
+        mode: Option<TimestampMode>,
+    ) -> Result<Vec<TranscriptionResult>> {
+        let mut results = Vec::with_capacity(audio_paths.len());
+        for path in audio_paths {
+            let result = self.transcribe_file(path, mode)?;
+            results.push(result);
+        }
+        Ok(results)
+    }
+
+    /// Get model directory path
+    pub fn model_dir(&self) -> &Path {
+        &self.model_dir
+    }
+}
diff --git a/vendor/parakeet-rs/src/sortformer.rs b/vendor/parakeet-rs/src/sortformer.rs
new file mode 100644
index 0000000..2b1e5a3
--- /dev/null
+++ b/vendor/parakeet-rs/src/sortformer.rs
@@ -0,0 +1,1062 @@
+//! NVIDIA Sortformer v2 Streaming Speaker Diarization
+//!
+//! This module implements NVIDIA's Sortformer v2 streaming model for speaker diarization.
+//!
+//! Key features:
+//! - Streaming inference with ~10s chunks (124 frames at 80ms each)
+//! - FIFO buffer for context management
+//! - Smart speaker cache compression (keeps important frames, not just recent)
+//! - Silence profile tracking
+//! - Post-processing: median filtering, hysteresis thresholding
+//! - Supports up to 4 speakers
+//!
+//! Reference: https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2
+//! Note that, my ONNX export:
+//! CHUNK_LEN = 124
+//! FIFO_LEN = 124
+//! CACHE_LEN = 188
+//! FEAT_DIM = 128
+//! EMB_DIM = 512
+//! Note, my stft code is adapted from: https://librosa.org/doc/main/generated/librosa.stft.html
+
+use crate::error::{Error, Result};
+use crate::execution::ModelConfig;
+use ndarray::{s, Array1, Array2, Array3, Axis};
+use ort::session::Session;
+use rustfft::{num_complex::Complex, FftPlanner};
+use std::f32::consts::PI;
+use std::path::Path;
+
+// Model constants
+const N_FFT: usize = 512;
+const WIN_LENGTH: usize = 400;
+const HOP_LENGTH: usize = 160;
+const N_MELS: usize = 128;
+const PREEMPH: f32 = 0.97;
+const LOG_ZERO_GUARD: f32 = 5.960464478e-8;
+const SAMPLE_RATE: usize = 16000;
+const FMIN: f32 = 0.0;
+const FMAX: f32 = 8000.0;
+
+// Streaming constants
+const CHUNK_LEN: usize = 124;           // Frames per chunk (~10s at 80ms)
+const FIFO_LEN: usize = 124;            // FIFO buffer length
+const SPKCACHE_LEN: usize = 188;        // Speaker cache length
+const SPKCACHE_UPDATE_PERIOD: usize = 124;
+const SUBSAMPLING: usize = 8;           // Audio frames -> model frames
+const EMB_DIM: usize = 512;             // Embedding dimension
+const NUM_SPEAKERS: usize = 4;          // Model supports 4 speakers
+const FRAME_DURATION: f32 = 0.08;       // 80ms per frame
+
+// Cache compression params (from NeMo)
+const SPKCACHE_SIL_FRAMES_PER_SPK: usize = 3;
+const PRED_SCORE_THRESHOLD: f32 = 0.25;
+const STRONG_BOOST_RATE: f32 = 0.75;
+const WEAK_BOOST_RATE: f32 = 1.5;
+const MIN_POS_SCORES_RATE: f32 = 0.5;
+const SIL_THRESHOLD: f32 = 0.2;
+const MAX_INDEX: usize = 99999;
+
+/// Post-processing configuration for speaker diarization. (NVIDIA official configs from v2 YAMLs)
+///
+/// Controls how raw model predictions are converted into speaker segments.
+/// NVIDIA provides pre-tuned configs for different datasets (CallHome, DIHARD3, AMI).
+///
+/// # Parameters
+/// - `onset`: Probability threshold to START a speaker segment (higher = more strict)
+/// - `offset`: Probability threshold to END a speaker segment (lower = longer segments)
+/// - `pad_onset`: Seconds to subtract from segment start times
+/// - `pad_offset`: Seconds to add to segment end times
+/// - `min_duration_on`: Minimum segment length in seconds (filters short blips)
+/// - `min_duration_off`: Minimum gap between segments before merging
+/// - `median_window`: Smoothing window size (odd number, higher = smoother)
+///
+/// # Pre-tuned Configs
+/// - `callhome()` - (default)
+/// - `dihard3()`
+///
+/// # Custom Config
+/// Use `custom(onset, offset)` to create your own config for fine-tuning.
+///
+/// See: https://github.com/NVIDIA-NeMo/NeMo/tree/main/examples/speaker_tasks/diarization/conf/neural_diarizer
+#[derive(Debug, Clone)]
+pub struct DiarizationConfig {
+    pub onset: f32,
+    pub offset: f32,
+    pub pad_onset: f32,
+    pub pad_offset: f32,
+    pub min_duration_on: f32,
+    pub min_duration_off: f32,
+    pub median_window: usize,
+}
+
+impl Default for DiarizationConfig {
+    fn default() -> Self {
+        Self::callhome()
+    }
+}
+
+impl DiarizationConfig {
+    /// CallHome dataset config for v2 (default)
+    /// From: diar_streaming_sortformer_4spk-v2_callhome-part1.yaml
+    pub fn callhome() -> Self {
+        Self {
+            onset: 0.641,
+            offset: 0.561,
+            pad_onset: 0.229,
+            pad_offset: 0.079,
+            min_duration_on: 0.511,
+            min_duration_off: 0.296,
+            median_window: 11,
+        }
+    }
+
+    /// DIHARD3 dataset config for v2
+    /// From: diar_streaming_sortformer_4spk-v2_dihard3-dev.yaml
+    pub fn dihard3() -> Self {
+        Self {
+            onset: 0.56,
+            offset: 1.0,
+            pad_onset: 0.063,
+            pad_offset: 0.002,
+            min_duration_on: 0.007,
+            min_duration_off: 0.151,
+            median_window: 11,
+        }
+    }
+
+    /// Create a custom config for fine-tuning diarization behavior.
+    ///
+    /// # Arguments
+    /// * `onset` - Probability threshold to start a segment (0.0-1.0, typical: 0.5-0.7)
+    /// * `offset` - Probability threshold to end a segment (0.0-1.0, typical: 0.4-0.6)
+    ///
+    /// # Example
+    /// ```rust
+    /// use parakeet_rs::sortformer::DiarizationConfig;
+    ///
+    /// // More sensitive detection (lower thresholds)
+    /// let sensitive = DiarizationConfig::custom(0.5, 0.4);
+    ///
+    /// // Stricter detection (higher thresholds, fewer false positives)
+    /// let strict = DiarizationConfig::custom(0.7, 0.6);
+    ///
+    /// // Full customization
+    /// let mut config = DiarizationConfig::custom(0.6, 0.5);
+    /// config.min_duration_on = 0.3;  // Ignore segments shorter than 300ms
+    /// config.median_window = 15;      // More smoothing
+    /// ```
+    pub fn custom(onset: f32, offset: f32) -> Self {
+        Self {
+            onset,
+            offset,
+            pad_onset: 0.0,
+            pad_offset: 0.0,
+            min_duration_on: 0.1,
+            min_duration_off: 0.1,
+            median_window: 11,
+        }
+    }
+}
+
+/// Speaker segment with start time, end time, and speaker ID
+#[derive(Debug, Clone)]
+pub struct SpeakerSegment {
+    pub start: f32,
+    pub end: f32,
+    pub speaker_id: usize,
+}
+
+/// Streaming Sortformer v2 speaker diarization engine
+pub struct Sortformer {
+    session: Session,
+    config: DiarizationConfig,
+    // Streaming state. note that, Same way as Nemo
+    spkcache: Array3<f32>,           // (1, 0..SPKCACHE_LEN, EMB_DIM)
+    spkcache_preds: Option<Array3<f32>>, // (1, 0..SPKCACHE_LEN, NUM_SPEAKERS)
+    fifo: Array3<f32>,               // (1, 0..FIFO_LEN, EMB_DIM)
+    fifo_preds: Array3<f32>,         // (1, 0..FIFO_LEN, NUM_SPEAKERS)
+    mean_sil_emb: Array2<f32>,       // (1, EMB_DIM)
+    n_sil_frames: usize,
+    // Mel filterbank (cached)
+    mel_basis: Array2<f32>,
+}
+
+impl Sortformer {
+    /// a new Sortformer instance from ONNX model path
+    pub fn new<P: AsRef<Path>>(model_path: P) -> Result<Self> {
+        Self::with_config(model_path, None, DiarizationConfig::default())
+    }
+
+    /// Create with custom config
+    pub fn with_config<P: AsRef<Path>>(
+        model_path: P,
+        execution_config: Option<ModelConfig>,
+        config: DiarizationConfig,
+    ) -> Result<Self> {
+        let config_to_use = execution_config.unwrap_or_default();
+
+        let session = config_to_use
+            .apply_to_session_builder(Session::builder()?)?
+            .commit_from_file(model_path.as_ref())?;
+
+        let mel_basis = Self::create_mel_filterbank();
+
+        let mut instance = Self {
+            session,
+            config,
+            spkcache: Array3::zeros((1, 0, EMB_DIM)),
+            spkcache_preds: None,
+            fifo: Array3::zeros((1, 0, EMB_DIM)),
+            fifo_preds: Array3::zeros((1, 0, NUM_SPEAKERS)),
+            mean_sil_emb: Array2::zeros((1, EMB_DIM)),
+            n_sil_frames: 0,
+            mel_basis,
+        };
+        instance.reset_state();
+        Ok(instance)
+    }
+
+    /// Reset streaming state
+    pub fn reset_state(&mut self) {
+        self.spkcache = Array3::zeros((1, 0, EMB_DIM));
+        self.spkcache_preds = None;
+        self.fifo = Array3::zeros((1, 0, EMB_DIM));
+        self.fifo_preds = Array3::zeros((1, 0, NUM_SPEAKERS));
+        self.mean_sil_emb = Array2::zeros((1, EMB_DIM));
+        self.n_sil_frames = 0;
+    }
+
+    /// Main diarization entry point
+    pub fn diarize(
+        &mut self,
+        mut audio: Vec<f32>,
+        sample_rate: u32,
+        channels: u16,
+    ) -> Result<Vec<SpeakerSegment>> {
+        // Resample if needed
+        if sample_rate != SAMPLE_RATE as u32 {
+            return Err(Error::Audio(format!(
+                "Expected {} Hz, got {} Hz",
+                SAMPLE_RATE, sample_rate
+            )));
+        }
+
+        // Convert to mono
+        if channels > 1 {
+            audio = audio
+                .chunks(channels as usize)
+                .map(|chunk| chunk.iter().sum::<f32>() / channels as f32)
+                .collect();
+        }
+
+        // Reset state for new audio
+        self.reset_state();
+
+        // Extract mel features (B, T, D)
+        let features = self.extract_mel_features(&audio);
+        let total_frames = features.shape()[1];
+
+        // Process in chunks
+        let chunk_stride = CHUNK_LEN * SUBSAMPLING;
+        let num_chunks = (total_frames + chunk_stride - 1) / chunk_stride;
+
+        let mut all_chunk_preds = Vec::new();
+
+        for chunk_idx in 0..num_chunks {
+            let start = chunk_idx * chunk_stride;
+            let end = (start + chunk_stride).min(total_frames);
+            let current_len = end - start;
+
+            // Extract chunk features
+            let mut chunk_feat = features.slice(s![.., start..end, ..]).to_owned();
+
+            // Pad last chunk if needed
+            if current_len < chunk_stride {
+                let mut padded = Array3::zeros((1, chunk_stride, N_MELS));
+                padded.slice_mut(s![.., ..current_len, ..]).assign(&chunk_feat);
+                chunk_feat = padded;
+            }
+
+            // Run streaming update
+            let chunk_preds = self.streaming_update(&chunk_feat, current_len)?;
+            all_chunk_preds.push(chunk_preds);
+        }
+
+        // Concatenate all predictions
+        let full_preds = Self::concat_predictions(&all_chunk_preds);
+
+        // Apply median filtering
+        let filtered_preds = if self.config.median_window > 1 {
+            self.median_filter(&full_preds)
+        } else {
+            full_preds
+        };
+
+        // Binarize to segments
+        let segments = self.binarize(&filtered_preds);
+
+        Ok(segments)
+    }
+
+    /// Streaming diarization that maintains state across calls.
+    ///
+    /// Unlike `diarize()`, this method does NOT reset the internal state,
+    /// allowing speaker embeddings to be preserved across multiple audio chunks.
+    /// Call `reset_state()` manually when starting a new audio session.
+    ///
+    /// This enables consistent speaker identification across long audio streams
+    /// by maintaining the speaker cache between processing windows.
+    ///
+    /// # Arguments
+    /// * `audio` - Audio samples (will be converted to mono if multi-channel)
+    /// * `sample_rate` - Must be 16000 Hz
+    /// * `channels` - Number of audio channels
+    ///
+    /// # Example
+    /// ```ignore
+    /// // Start of session
+    /// sortformer.reset_state();
+    ///
+    /// // Process sliding windows
+    /// let segments1 = sortformer.diarize_streaming(window1, 16000, 1)?;
+    /// let segments2 = sortformer.diarize_streaming(window2, 16000, 1)?; // Maintains speaker IDs
+    /// ```
+    pub fn diarize_streaming(
+        &mut self,
+        mut audio: Vec<f32>,
+        sample_rate: u32,
+        channels: u16,
+    ) -> Result<Vec<SpeakerSegment>> {
+        // Resample if needed
+        if sample_rate != SAMPLE_RATE as u32 {
+            return Err(Error::Audio(format!(
+                "Expected {} Hz, got {} Hz",
+                SAMPLE_RATE, sample_rate
+            )));
+        }
+
+        // Convert to mono
+        if channels > 1 {
+            audio = audio
+                .chunks(channels as usize)
+                .map(|chunk| chunk.iter().sum::<f32>() / channels as f32)
+                .collect();
+        }
+
+        // NOTE: Unlike diarize(), we do NOT call reset_state() here
+        // This preserves speaker embeddings across calls
+
+        // Extract mel features (B, T, D)
+        let features = self.extract_mel_features(&audio);
+        let total_frames = features.shape()[1];
+
+        // Process in chunks
+        let chunk_stride = CHUNK_LEN * SUBSAMPLING;
+        let num_chunks = (total_frames + chunk_stride - 1) / chunk_stride;
+
+        let mut all_chunk_preds = Vec::new();
+
+        for chunk_idx in 0..num_chunks {
+            let start = chunk_idx * chunk_stride;
+            let end = (start + chunk_stride).min(total_frames);
+            let current_len = end - start;
+
+            // Extract chunk features
+            let mut chunk_feat = features.slice(s![.., start..end, ..]).to_owned();
+
+            // Pad last chunk if needed
+            if current_len < chunk_stride {
+                let mut padded = Array3::zeros((1, chunk_stride, N_MELS));
+                padded.slice_mut(s![.., ..current_len, ..]).assign(&chunk_feat);
+                chunk_feat = padded;
+            }
+
+            // Run streaming update
+            let chunk_preds = self.streaming_update(&chunk_feat, current_len)?;
+            all_chunk_preds.push(chunk_preds);
+        }
+
+        // Concatenate all predictions
+        let full_preds = Self::concat_predictions(&all_chunk_preds);
+
+        // Apply median filtering
+        let filtered_preds = if self.config.median_window > 1 {
+            self.median_filter(&full_preds)
+        } else {
+            full_preds
+        };
+
+        // Binarize to segments
+        let segments = self.binarize(&filtered_preds);
+
+        Ok(segments)
+    }
+
+    /// NeMo's streaming_update with smart cache compression.
+    /// Public to allow incremental streaming diarization.
+    pub fn streaming_update(&mut self, chunk_feat: &Array3<f32>, current_len: usize) -> Result<Array2<f32>> {
+        let spkcache_len = self.spkcache.shape()[1];
+        let fifo_len = self.fifo.shape()[1];
+
+        // Prepare inputs
+        let chunk_lengths = Array1::from_vec(vec![current_len as i64]);
+        let spkcache_lengths = Array1::from_vec(vec![spkcache_len as i64]);
+        let fifo_lengths = Array1::from_vec(vec![fifo_len as i64]);
+
+        // Prepare FIFO input
+        let fifo_input = if fifo_len > 0 {
+            self.fifo.clone()
+        } else {
+            Array3::zeros((1, 0, EMB_DIM))
+        };
+
+        // Prepare spkcache input (may be empty)
+        let spkcache_input = if spkcache_len > 0 {
+            self.spkcache.clone()
+        } else {
+            Array3::zeros((1, 0, EMB_DIM))
+        };
+
+        // Create input values
+        let chunk_value = ort::value::Value::from_array(chunk_feat.clone())?;
+        let chunk_lengths_value = ort::value::Value::from_array(chunk_lengths)?;
+        let spkcache_value = ort::value::Value::from_array(spkcache_input)?;
+        let spkcache_lengths_value = ort::value::Value::from_array(spkcache_lengths)?;
+        let fifo_value = ort::value::Value::from_array(fifo_input)?;
+        let fifo_lengths_value = ort::value::Value::from_array(fifo_lengths)?;
+
+        // Run ONNX inference and extract all data in a block to release borrow
+        let (preds, new_embs, chunk_len) = {
+            let outputs = self.session.run(ort::inputs!(
+                "chunk" => chunk_value,
+                "chunk_lengths" => chunk_lengths_value,
+                "spkcache" => spkcache_value,
+                "spkcache_lengths" => spkcache_lengths_value,
+                "fifo" => fifo_value,
+                "fifo_lengths" => fifo_lengths_value
+            ))?;
+
+            // Extract outputs
+            let (preds_shape, preds_data) = outputs["spkcache_fifo_chunk_preds"]
+                .try_extract_tensor::<f32>()
+                .map_err(|e| Error::Model(format!("Failed to extract preds: {e}")))?;
+            let (embs_shape, embs_data) = outputs["chunk_pre_encode_embs"]
+                .try_extract_tensor::<f32>()
+                .map_err(|e| Error::Model(format!("Failed to extract embs: {e}")))?;
+
+            // Convert to ndarray
+            let preds_dims = preds_shape.as_ref();
+            let embs_dims = embs_shape.as_ref();
+
+            let preds = Array3::from_shape_vec(
+                (preds_dims[0] as usize, preds_dims[1] as usize, preds_dims[2] as usize),
+                preds_data.to_vec()
+            ).map_err(|e| Error::Model(format!("Failed to reshape preds: {e}")))?;
+
+            let new_embs = Array3::from_shape_vec(
+                (embs_dims[0] as usize, embs_dims[1] as usize, embs_dims[2] as usize),
+                embs_data.to_vec()
+            ).map_err(|e| Error::Model(format!("Failed to reshape embs: {e}")))?;
+
+            // Calculate valid frames
+            let valid_frames = (current_len + SUBSAMPLING - 1) / SUBSAMPLING;
+
+            (preds, new_embs, valid_frames)
+        };
+
+        // Extract predictions for different parts
+        let fifo_preds = if fifo_len > 0 {
+            preds.slice(s![0, spkcache_len..spkcache_len + fifo_len, ..]).to_owned()
+        } else {
+            Array2::zeros((0, NUM_SPEAKERS))
+        };
+
+        let chunk_preds = preds.slice(s![0, spkcache_len + fifo_len..spkcache_len + fifo_len + chunk_len, ..]).to_owned();
+        let chunk_embs = new_embs.slice(s![0, ..chunk_len, ..]).to_owned();
+
+        // Append chunk embeddings to FIFO
+        self.fifo = Self::concat_axis1(&self.fifo, &chunk_embs.insert_axis(Axis(0)));
+
+        // Update FIFO predictions
+        if fifo_len > 0 {
+            let combined = Self::concat_axis1_2d(&fifo_preds, &chunk_preds);
+            self.fifo_preds = combined.insert_axis(Axis(0));
+        } else {
+            self.fifo_preds = chunk_preds.clone().insert_axis(Axis(0));
+        }
+
+        let fifo_len_after = self.fifo.shape()[1];
+
+        // Move from FIFO to cache when FIFO exceeds limit
+        if fifo_len_after > FIFO_LEN {
+            let mut pop_out_len = SPKCACHE_UPDATE_PERIOD;
+            pop_out_len = pop_out_len.max(chunk_len.saturating_sub(FIFO_LEN) + fifo_len);
+            pop_out_len = pop_out_len.min(fifo_len_after);
+
+            let pop_out_embs = self.fifo.slice(s![.., ..pop_out_len, ..]).to_owned();
+            let pop_out_preds = self.fifo_preds.slice(s![.., ..pop_out_len, ..]).to_owned();
+
+            // Update silence profile
+            self.update_silence_profile(&pop_out_embs, &pop_out_preds);
+
+            // Remove from FIFO
+            self.fifo = self.fifo.slice(s![.., pop_out_len.., ..]).to_owned();
+            self.fifo_preds = self.fifo_preds.slice(s![.., pop_out_len.., ..]).to_owned();
+
+            // Append to cache
+            self.spkcache = Self::concat_axis1(&self.spkcache, &pop_out_embs);
+
+            if let Some(ref cache_preds) = self.spkcache_preds {
+                self.spkcache_preds = Some(Self::concat_axis1(cache_preds, &pop_out_preds));
+            }
+
+            // Smart compression when cache exceeds limit
+            if self.spkcache.shape()[1] > SPKCACHE_LEN {
+                if self.spkcache_preds.is_none() {
+                    // Initialize cache predictions from initial output
+                    let initial_cache_preds = preds.slice(s![.., ..spkcache_len, ..]).to_owned();
+                    let combined = Self::concat_axis1(&initial_cache_preds, &pop_out_preds);
+                    self.spkcache_preds = Some(combined);
+                }
+
+                // Use smart compression
+                self.compress_spkcache();
+            }
+        }
+
+        Ok(chunk_preds)
+    }
+
+    /// Update mean silence embedding
+    fn update_silence_profile(&mut self, embs: &Array3<f32>, preds: &Array3<f32>) {
+        let preds_2d = preds.slice(s![0, .., ..]);
+
+        for t in 0..preds_2d.shape()[0] {
+            let sum: f32 = (0..NUM_SPEAKERS).map(|s| preds_2d[[t, s]]).sum();
+            if sum < SIL_THRESHOLD {
+                // This is a silence frame
+                let emb = embs.slice(s![0, t, ..]);
+
+                // Update running mean
+                let old_sum: Vec<f32> = self.mean_sil_emb.slice(s![0, ..]).iter()
+                    .map(|&x| x * self.n_sil_frames as f32)
+                    .collect();
+
+                self.n_sil_frames += 1;
+
+                for i in 0..EMB_DIM {
+                    self.mean_sil_emb[[0, i]] = (old_sum[i] + emb[i]) / self.n_sil_frames as f32;
+                }
+            }
+        }
+    }
+
+    /// Smart cache compression
+    fn compress_spkcache(&mut self) {
+        let cache_preds = match &self.spkcache_preds {
+            Some(p) => p.clone(),
+            None => return,
+        };
+
+        let n_frames = self.spkcache.shape()[1];
+        let spkcache_len_per_spk = SPKCACHE_LEN / NUM_SPEAKERS - SPKCACHE_SIL_FRAMES_PER_SPK;
+        let strong_boost_per_spk = (spkcache_len_per_spk as f32 * STRONG_BOOST_RATE) as usize;
+        let weak_boost_per_spk = (spkcache_len_per_spk as f32 * WEAK_BOOST_RATE) as usize;
+        let min_pos_scores_per_spk = (spkcache_len_per_spk as f32 * MIN_POS_SCORES_RATE) as usize;
+
+        // Calculate quality scores
+        let preds_2d = cache_preds.slice(s![0, .., ..]).to_owned();
+        let mut scores = self.get_log_pred_scores(&preds_2d);
+
+        // Disable low scores
+        scores = self.disable_low_scores(&preds_2d, scores, min_pos_scores_per_spk);
+
+        // Boost important frames
+        scores = self.boost_topk_scores(scores, strong_boost_per_spk, 2.0);
+        scores = self.boost_topk_scores(scores, weak_boost_per_spk, 1.0);
+
+        // Add silence frames placeholder
+        if SPKCACHE_SIL_FRAMES_PER_SPK > 0 {
+            let mut padded = Array2::from_elem((n_frames + SPKCACHE_SIL_FRAMES_PER_SPK, NUM_SPEAKERS), f32::NEG_INFINITY);
+            padded.slice_mut(s![..n_frames, ..]).assign(&scores);
+            for i in n_frames..n_frames + SPKCACHE_SIL_FRAMES_PER_SPK {
+                for j in 0..NUM_SPEAKERS {
+                    padded[[i, j]] = f32::INFINITY;
+                }
+            }
+            scores = padded;
+        }
+
+        // Select top frames
+        let (topk_indices, is_disabled) = self.get_topk_indices(&scores, n_frames);
+
+        // Gather embeddings
+        let (new_embs, new_preds) = self.gather_spkcache(&topk_indices, &is_disabled);
+
+        self.spkcache = new_embs;
+        self.spkcache_preds = Some(new_preds);
+    }
+
+    /// Calculate quality scores
+    fn get_log_pred_scores(&self, preds: &Array2<f32>) -> Array2<f32> {
+        let mut scores = Array2::zeros(preds.dim());
+
+        for t in 0..preds.shape()[0] {
+            let mut log_1_probs_sum = 0.0f32;
+            for s in 0..NUM_SPEAKERS {
+                let p = preds[[t, s]].max(PRED_SCORE_THRESHOLD);
+                let log_1_p = (1.0 - p).max(PRED_SCORE_THRESHOLD).ln();
+                log_1_probs_sum += log_1_p;
+            }
+
+            for s in 0..NUM_SPEAKERS {
+                let p = preds[[t, s]].max(PRED_SCORE_THRESHOLD);
+                let log_p = p.ln();
+                let log_1_p = (1.0 - p).max(PRED_SCORE_THRESHOLD).ln();
+                scores[[t, s]] = log_p - log_1_p + log_1_probs_sum - 0.5f32.ln();
+            }
+        }
+
+        scores
+    }
+
+    /// Disable non-speech and overlapped speech
+    fn disable_low_scores(&self, preds: &Array2<f32>, mut scores: Array2<f32>, min_pos_scores_per_spk: usize) -> Array2<f32> {
+        // Count positive scores per speaker
+        let mut pos_count = vec![0usize; NUM_SPEAKERS];
+        for t in 0..scores.shape()[0] {
+            for s in 0..NUM_SPEAKERS {
+                if scores[[t, s]] > 0.0 {
+                    pos_count[s] += 1;
+                }
+            }
+        }
+
+        for t in 0..preds.shape()[0] {
+            for s in 0..NUM_SPEAKERS {
+                let is_speech = preds[[t, s]] > 0.5;
+
+                if !is_speech {
+                    scores[[t, s]] = f32::NEG_INFINITY;
+                } else {
+                    let is_pos = scores[[t, s]] > 0.0;
+                    if !is_pos && pos_count[s] >= min_pos_scores_per_spk {
+                        scores[[t, s]] = f32::NEG_INFINITY;
+                    }
+                }
+            }
+        }
+
+        scores
+    }
+
+    /// Boost top K frames per speaker
+    fn boost_topk_scores(&self, mut scores: Array2<f32>, n_boost_per_spk: usize, scale_factor: f32) -> Array2<f32> {
+        for s in 0..NUM_SPEAKERS {
+            // Get column for this speaker
+            let col: Vec<(usize, f32)> = (0..scores.shape()[0])
+                .map(|t| (t, scores[[t, s]]))
+                .collect();
+
+            // Sort by score descending
+            let mut sorted = col.clone();
+            sorted.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
+
+            // Boost top K
+            for i in 0..n_boost_per_spk.min(sorted.len()) {
+                let t = sorted[i].0;
+                if scores[[t, s]] != f32::NEG_INFINITY {
+                    scores[[t, s]] -= scale_factor * 0.5f32.ln();
+                }
+            }
+        }
+
+        scores
+    }
+
+    /// Get indices of top frames
+    fn get_topk_indices(&self, scores: &Array2<f32>, n_frames_no_sil: usize) -> (Vec<usize>, Vec<bool>) {
+        let n_frames = scores.shape()[0];
+
+        // Flatten scores as (S, T) then reshape to (S*T,)
+        // This means we iterate: speaker 0 all times, then speaker 1 all times, etc.
+        // flat_index = speaker * n_frames + time
+        let mut flat_scores: Vec<(usize, f32)> = Vec::with_capacity(n_frames * NUM_SPEAKERS);
+        for s in 0..NUM_SPEAKERS {
+            for t in 0..n_frames {
+                let flat_idx = s * n_frames + t;
+                flat_scores.push((flat_idx, scores[[t, s]]));
+            }
+        }
+
+        // Sort by score descending to get top-K
+        flat_scores.sort_by(|a, b| {
+            b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal)
+        });
+
+        // Take top SPKCACHE_LEN and replace invalid scores with MAX_INDEX
+        let mut topk_flat: Vec<usize> = flat_scores
+            .iter()
+            .take(SPKCACHE_LEN)
+            .map(|(idx, score)| {
+                if *score == f32::NEG_INFINITY {
+                    MAX_INDEX
+                } else {
+                    *idx
+                }
+            })
+            .collect();
+
+        // Sort flat indices ascending (this puts MAX_INDEX at the end)
+        topk_flat.sort();
+
+        // Compute is_disabled and convert to frame indices
+        let mut is_disabled = vec![false; SPKCACHE_LEN];
+        let mut frame_indices = vec![0usize; SPKCACHE_LEN];
+
+        for (i, &flat_idx) in topk_flat.iter().enumerate() {
+            if flat_idx == MAX_INDEX {
+                // Invalid entries are disabled
+                is_disabled[i] = true;
+                frame_indices[i] = 0;  // We set disabled to 0
+            } else {
+                // convert to frame index
+                let frame_idx = flat_idx % n_frames;
+                
+                // check if frame is beyond valid range
+                if frame_idx >= n_frames_no_sil {
+                    is_disabled[i] = true;
+                    frame_indices[i] = 0;  // same as abov: set disabled to 0
+                } else {
+                    frame_indices[i] = frame_idx;
+                }
+            }
+        }
+
+        (frame_indices, is_disabled)
+    }
+
+    /// Gather selected frames
+    fn gather_spkcache(&self, indices: &[usize], is_disabled: &[bool]) -> (Array3<f32>, Array3<f32>) {
+        let mut new_embs = Array3::zeros((1, SPKCACHE_LEN, EMB_DIM));
+        let mut new_preds = Array3::zeros((1, SPKCACHE_LEN, NUM_SPEAKERS));
+
+        let cache_preds = self.spkcache_preds.as_ref().unwrap();
+
+        for (i, (&idx, &disabled)) in indices.iter().zip(is_disabled.iter()).enumerate() {
+            if i >= SPKCACHE_LEN {
+                break;
+            }
+
+            if disabled {
+                // Use silence embedding
+                new_embs.slice_mut(s![0, i, ..]).assign(&self.mean_sil_emb.slice(s![0, ..]));
+                // Predictions stay zero
+            } else if idx < self.spkcache.shape()[1] {
+                new_embs.slice_mut(s![0, i, ..]).assign(&self.spkcache.slice(s![0, idx, ..]));
+                new_preds.slice_mut(s![0, i, ..]).assign(&cache_preds.slice(s![0, idx, ..]));
+            }
+        }
+
+        (new_embs, new_preds)
+    }
+
+    /// Concatenate along axis 1 for 3D arrays
+    fn concat_axis1(a: &Array3<f32>, b: &Array3<f32>) -> Array3<f32> {
+        if a.shape()[1] == 0 {
+            return b.clone();
+        }
+        if b.shape()[1] == 0 {
+            return a.clone();
+        }
+        ndarray::concatenate(Axis(1), &[a.view(), b.view()]).unwrap()
+    }
+
+    /// Concatenate along axis 0 for 2D arrays
+    fn concat_axis1_2d(a: &Array2<f32>, b: &Array2<f32>) -> Array2<f32> {
+        if a.shape()[0] == 0 {
+            return b.clone();
+        }
+        if b.shape()[0] == 0 {
+            return a.clone();
+        }
+        ndarray::concatenate(Axis(0), &[a.view(), b.view()]).unwrap()
+    }
+
+    /// Concatenate predictions
+    fn concat_predictions(preds: &[Array2<f32>]) -> Array2<f32> {
+        if preds.is_empty() {
+            return Array2::zeros((0, NUM_SPEAKERS));
+        }
+        if preds.len() == 1 {
+            return preds[0].clone();
+        }
+
+        let views: Vec<_> = preds.iter().map(|p| p.view()).collect();
+        ndarray::concatenate(Axis(0), &views).unwrap()
+    }
+
+    /// Apply median filter to predictions
+    fn median_filter(&self, preds: &Array2<f32>) -> Array2<f32> {
+        let window = self.config.median_window;
+        let half = window / 2;
+        let mut filtered = preds.clone();
+
+        for spk in 0..NUM_SPEAKERS {
+            for t in 0..preds.shape()[0] {
+                let start = t.saturating_sub(half);
+                let end = (t + half + 1).min(preds.shape()[0]);
+
+                let mut values: Vec<f32> = (start..end)
+                    .map(|i| preds[[i, spk]])
+                    .collect();
+                values.sort_by(|a, b| a.partial_cmp(b).unwrap());
+
+                filtered[[t, spk]] = values[values.len() / 2];
+            }
+        }
+
+        filtered
+    }
+
+    /// Binarize predictions to segments (padding applied during thresholding)
+    fn binarize(&self, preds: &Array2<f32>) -> Vec<SpeakerSegment> {
+        let mut segments = Vec::new();
+        let num_frames = preds.shape()[0];
+
+        for spk in 0..NUM_SPEAKERS {
+            let mut in_seg = false;
+            let mut seg_start = 0;
+            let mut temp_segments = Vec::new();
+
+            for t in 0..num_frames {
+                let p = preds[[t, spk]];
+
+                if p >= self.config.onset && !in_seg {
+                    in_seg = true;
+                    seg_start = t;
+                } else if p < self.config.offset && in_seg {
+                    in_seg = false;
+
+                    // Apply padding during conversion
+                    let start_t = (seg_start as f32 * FRAME_DURATION - self.config.pad_onset).max(0.0);
+                    let end_t = t as f32 * FRAME_DURATION + self.config.pad_offset;
+
+                    if end_t - start_t >= self.config.min_duration_on {
+                        temp_segments.push(SpeakerSegment {
+                            start: start_t,
+                            end: end_t,
+                            speaker_id: spk,
+                        });
+                    }
+                }
+            }
+
+            // Handle segment at end
+            if in_seg {
+                let start_t = (seg_start as f32 * FRAME_DURATION - self.config.pad_onset).max(0.0);
+                let end_t = num_frames as f32 * FRAME_DURATION + self.config.pad_offset;
+
+                if end_t - start_t >= self.config.min_duration_on {
+                    temp_segments.push(SpeakerSegment {
+                        start: start_t,
+                        end: end_t,
+                        speaker_id: spk,
+                    });
+                }
+            }
+
+            // Merge close segments (min_duration_off)
+            if temp_segments.len() > 1 {
+                let mut filtered = vec![temp_segments[0].clone()];
+                for seg in temp_segments.into_iter().skip(1) {
+                    let last = filtered.last_mut().unwrap();
+                    let gap = seg.start - last.end;
+                    if gap < self.config.min_duration_off {
+                        last.end = seg.end; // Merge
+                    } else {
+                        filtered.push(seg);
+                    }
+                }
+                segments.extend(filtered);
+            } else {
+                segments.extend(temp_segments);
+            }
+        }
+
+        // Sort by start time
+        segments.sort_by(|a, b| a.start.partial_cmp(&b.start).unwrap());
+        segments
+    }
+
+
+    fn apply_preemphasis(audio: &[f32]) -> Vec<f32> {
+        let mut result = Vec::with_capacity(audio.len());
+        result.push(audio[0]);
+        for i in 1..audio.len() {
+            result.push(audio[i] - PREEMPH * audio[i - 1]);
+        }
+        result
+    }
+
+    fn hann_window(window_length: usize) -> Vec<f32> {
+        // Librosa uses periodic window (fftbins=True): divide by N, not N-1
+        (0..window_length)
+            .map(|i| 0.5 - 0.5 * ((2.0 * PI * i as f32) / window_length as f32).cos())
+            .collect()
+    }
+
+    fn stft(audio: &[f32]) -> Array2<f32> {
+        let mut planner = FftPlanner::<f32>::new();
+        let fft = planner.plan_fft_forward(N_FFT);
+
+        // Create Hann window of length win_length, then zero-pad to n_fft (centered)
+        // This is exactly what librosa does: util.pad_center(fft_window, size=n_fft)
+        let hann = Self::hann_window(WIN_LENGTH);
+        let win_offset = (N_FFT - WIN_LENGTH) / 2;
+        let mut fft_window = vec![0.0f32; N_FFT];
+        for i in 0..WIN_LENGTH {
+            fft_window[win_offset + i] = hann[i];
+        }
+
+        // Pad signal for center=True (like librosa/torch.stft)
+        // Padding is n_fft // 2 on each side
+        let pad_amount = N_FFT / 2;
+        let mut padded_audio = vec![0.0; pad_amount];
+        padded_audio.extend_from_slice(audio);
+        padded_audio.extend(vec![0.0; pad_amount]);
+
+        let num_frames = (padded_audio.len() - N_FFT) / HOP_LENGTH + 1;
+        let freq_bins = N_FFT / 2 + 1;
+        let mut spectrogram = Array2::<f32>::zeros((freq_bins, num_frames));
+
+        for frame_idx in 0..num_frames {
+            let start = frame_idx * HOP_LENGTH;
+            let mut frame: Vec<Complex<f32>> = vec![Complex::new(0.0, 0.0); N_FFT];
+
+            // Extract n_fft samples and multiply by zero-padded window
+            for i in 0..N_FFT {
+                if start + i < padded_audio.len() {
+                    frame[i] = Complex::new(padded_audio[start + i] * fft_window[i], 0.0);
+                }
+            }
+
+            fft.process(&mut frame);
+            for k in 0..freq_bins {
+                let magnitude = frame[k].norm();
+                // Power spectrum (magnitude^2) - NeMo uses mag_power=2.0
+                spectrogram[[k, frame_idx]] = magnitude * magnitude;
+            }
+        }
+
+        spectrogram
+    }
+
+    // Librosa's Slaney mel scale (htk=False, which is the default)
+    fn hz_to_mel_slaney(hz: f64) -> f64 {
+        let f_min = 0.0;
+        let f_sp = 200.0 / 3.0;
+        let min_log_hz = 1000.0;
+        let min_log_mel = (min_log_hz - f_min) / f_sp;
+        let logstep = (6.4f64).ln() / 27.0;
+
+        if hz >= min_log_hz {
+            min_log_mel + (hz / min_log_hz).ln() / logstep
+        } else {
+            (hz - f_min) / f_sp
+        }
+    }
+
+    fn mel_to_hz_slaney(mel: f64) -> f64 {
+        let f_min = 0.0;
+        let f_sp = 200.0 / 3.0;
+        let min_log_hz = 1000.0;
+        let min_log_mel = (min_log_hz - f_min) / f_sp;
+        let logstep = (6.4f64).ln() / 27.0;
+
+        if mel >= min_log_mel {
+            min_log_hz * (logstep * (mel - min_log_mel)).exp()
+        } else {
+            f_min + f_sp * mel
+        }
+    }
+
+    fn create_mel_filterbank() -> Array2<f32> {
+        // lets use f64 for intermediate calculations to avoid precision loss
+        let freq_bins = N_FFT / 2 + 1;
+        let mut filterbank = Array2::<f32>::zeros((N_MELS, freq_bins));
+
+        // FFT frequencies: fftfreqs[k] = k * sr / n_fft
+        let fftfreqs: Vec<f64> = (0..freq_bins)
+            .map(|k| k as f64 * SAMPLE_RATE as f64 / N_FFT as f64)
+            .collect();
+
+        // Mel center frequencies using Slaney scale (librosa default, htk=False)
+        let fmin_mel = Self::hz_to_mel_slaney(FMIN as f64);
+        let fmax_mel = Self::hz_to_mel_slaney(FMAX as f64);
+        let mel_f: Vec<f64> = (0..=N_MELS + 1)
+            .map(|i| {
+                let mel = fmin_mel + (fmax_mel - fmin_mel) * i as f64 / (N_MELS + 1) as f64;
+                Self::mel_to_hz_slaney(mel)
+            })
+            .collect();
+
+        // Differences between consecutive mel frequencies
+        let fdiff: Vec<f64> = mel_f.windows(2).map(|w| w[1] - w[0]).collect();
+
+        // Compute filterbank weights (reference: librosa's ramp method)
+        // https://librosa.org/doc/main/generated/librosa.stft.html
+        for i in 0..N_MELS {
+            for k in 0..freq_bins {
+                // Lower slope: (fftfreqs[k] - mel_f[i]) / fdiff[i]
+                let lower = (fftfreqs[k] - mel_f[i]) / fdiff[i];
+                // Upper slope: (mel_f[i+2] - fftfreqs[k]) / fdiff[i+1]
+                let upper = (mel_f[i + 2] - fftfreqs[k]) / fdiff[i + 1];
+                // Weight is max(0, min(lower, upper))
+                filterbank[[i, k]] = 0.0f64.max(lower.min(upper)) as f32;
+            }
+        }
+
+        // Apply Slaney normalization: 2.0 / (mel_f[i+2] - mel_f[i])
+        for i in 0..N_MELS {
+            let enorm = 2.0 / (mel_f[i + 2] - mel_f[i]);
+            for k in 0..freq_bins {
+                filterbank[[i, k]] *= enorm as f32;
+            }
+        }
+
+        filterbank
+    }
+
+    fn extract_mel_features(&self, audio: &[f32]) -> Array3<f32> {
+        // 1. Add dither (small random noise to prevent log(0))
+        // NeMo uses dither=1e-5, but for determinism we skip random noise
+        // The log_zero_guard handles zero values
+
+        // 2. Apply preemphasis (NeMo uses preemph=0.97)
+        let preemphasized = Self::apply_preemphasis(audio);
+
+        // 3. STFT
+        let spectrogram = Self::stft(&preemphasized);
+
+        // 4. Apply mel filterbank (with Slaney normalization)
+        let mel_spec = self.mel_basis.dot(&spectrogram);
+
+        // 5. Log with guard value (NeMo uses log_zero_guard_value = 2^-24)
+        // NeMo uses normalize='NA' which means NO normalization
+        let log_mel_spec = mel_spec.mapv(|x| (x + LOG_ZERO_GUARD).ln());
+
+        let num_frames = log_mel_spec.shape()[1];
+        let mut features = Array3::<f32>::zeros((1, num_frames, N_MELS));
+
+        // Transpose to (batch, time, features) - NeMo outputs (B, D, T), model expects (B, T, D)
+        for t in 0..num_frames {
+            for m in 0..N_MELS {
+                features[[0, t, m]] = log_mel_spec[[m, t]];
+            }
+        }
+
+        features
+    }
+}
diff --git a/vendor/parakeet-rs/src/timestamps.rs b/vendor/parakeet-rs/src/timestamps.rs
new file mode 100644
index 0000000..81ea600
--- /dev/null
+++ b/vendor/parakeet-rs/src/timestamps.rs
@@ -0,0 +1,280 @@
+use crate::decoder::TimedToken;
+
+/// Timestamp output mode for transcription results
+///
+/// Determines how token-level timestamps are grouped and presented:
+/// - `Tokens`: Raw token-level output from the model (most detailed)
+/// - `Words`: Tokens grouped into individual words
+/// - `Sentences`: Tokens grouped by sentence boundaries (., ?, !)
+///
+/// # Model-Specific Recommendations
+///
+/// - **Parakeet CTC (English)**: Use `Words` mode. The CTC model only outputs lowercase
+///   alphabet without punctuation, so sentence segmentation is not possible.
+/// - **Parakeet TDT (Multilingual)**: Use `Sentences` mode. The TDT model predicts
+///   punctuation, enabling natural sentence boundaries.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum TimestampMode {
+    /// Raw token-level timestamps from the model
+    Tokens,
+    /// Word-level timestamps (groups subword tokens)
+    Words,
+    /// Sentence-level timestamps (groups by punctuation)
+    ///
+    /// Note: Only works with models that predict punctuation (e.g., Parakeet TDT).
+    /// CTC models don't predict punctuation, so use `Words` mode instead.
+    Sentences,
+}
+
+impl Default for TimestampMode {
+    fn default() -> Self {
+        Self::Tokens
+    }
+}
+
+/// Convert token timestamps to the requested output mode
+///
+/// Takes raw token-level timestamps from the model and optionally groups them
+/// into words or sentences while preserving the original timing information.
+///
+/// # Arguments
+///
+/// * `tokens` - Raw token-level timestamps from model output
+/// * `mode` - Desired grouping level (Tokens, Words, or Sentences)
+///
+/// # Returns
+///
+/// Vector of TimedToken with timestamps at the requested granularity
+pub fn process_timestamps(tokens: &[TimedToken], mode: TimestampMode) -> Vec<TimedToken> {
+    match mode {
+        TimestampMode::Tokens => tokens.to_vec(),
+        TimestampMode::Words => group_by_words(tokens),
+        TimestampMode::Sentences => group_by_sentences(tokens),
+    }
+}
+
+// Group tokens into words based on word boundary markers
+fn group_by_words(tokens: &[TimedToken]) -> Vec<TimedToken> {
+    if tokens.is_empty() {
+        return Vec::new();
+    }
+
+    let mut words = Vec::new();
+    let mut current_word_text = String::new();
+    let mut current_word_start = 0.0;
+    let mut last_word_lower = String::new();
+
+    for (i, token) in tokens.iter().enumerate() {
+        // Skip empty tokens
+        if token.text.trim().is_empty() {
+            continue;
+        }
+
+        // Check if this starts a new word (SentencePiece uses ▁ or space prefix)
+        // Also treat PURE punctuation marks (like ".", ",") as separate words
+        // But NOT contractions like "'re" or "'s" which should attach to previous word
+        let is_pure_punctuation = !token.text.is_empty() &&
+            token.text.chars().all(|c| c.is_ascii_punctuation());
+
+        // Check if this is a contraction suffix
+        // These should NOT start a new word - they attach to the previous word
+        let token_without_marker = token.text.trim_start_matches('▁').trim_start_matches(' ');
+        let is_contraction = token_without_marker.starts_with('\'');
+
+        let starts_word = (token.text.starts_with('▁')
+            || token.text.starts_with(' ')
+            || is_pure_punctuation)
+            && !is_contraction
+            || i == 0;
+
+        if starts_word && !current_word_text.is_empty() {
+            // Save previous word (with deduplication)
+            let word_lower = current_word_text.to_lowercase();
+            if word_lower != last_word_lower {
+                words.push(TimedToken {
+                    text: current_word_text.clone(),
+                    start: current_word_start,
+                    end: tokens[i - 1].end,
+                });
+                last_word_lower = word_lower;
+            }
+            current_word_text.clear();
+        }
+
+        // Start new word or append to current
+        if current_word_text.is_empty() {
+            current_word_start = token.start;
+        }
+
+        // Add token text, removing word boundary markers
+        let token_text = token
+            .text
+            .trim_start_matches('▁')
+            .trim_start_matches(' ');
+        current_word_text.push_str(token_text);
+    }
+
+    // Add final word
+    if !current_word_text.is_empty() {
+        let word_lower = current_word_text.to_lowercase();
+        if word_lower != last_word_lower {
+            words.push(TimedToken {
+                text: current_word_text,
+                start: current_word_start,
+                end: tokens.last().unwrap().end,
+            });
+        }
+    }
+
+    words
+}
+
+// Group words into sentences based on punctuation
+fn group_by_sentences(tokens: &[TimedToken]) -> Vec<TimedToken> {
+    // First get word-level grouping
+    let words = group_by_words(tokens);
+    if words.is_empty() {
+        return Vec::new();
+    }
+
+    let mut sentences = Vec::new();
+    let mut current_sentence = Vec::new();
+
+    for word in words {
+        current_sentence.push(word.clone());
+
+        // Check if word ends with sentence terminator
+        let ends_sentence = word.text.contains('.')
+            || word.text.contains('?')
+            || word.text.contains('!');
+
+        if ends_sentence {
+            let sentence_text = format_sentence(&current_sentence);
+            let start = current_sentence.first().unwrap().start;
+            let end = current_sentence.last().unwrap().end;
+
+            if !sentence_text.is_empty() {
+                sentences.push(TimedToken {
+                    text: sentence_text,
+                    start,
+                    end,
+                });
+            }
+            current_sentence.clear();
+        }
+    }
+
+    // Add final sentence if exists
+    if !current_sentence.is_empty() {
+        let sentence_text = format_sentence(&current_sentence);
+        let start = current_sentence.first().unwrap().start;
+        let end = current_sentence.last().unwrap().end;
+
+        if !sentence_text.is_empty() {
+            sentences.push(TimedToken {
+                text: sentence_text,
+                start,
+                end,
+            });
+        }
+    }
+
+    sentences
+}
+
+// Join words with punctuation spacing
+fn format_sentence(words: &[TimedToken]) -> String {
+    let result: Vec<&str> = words.iter().map(|w| w.text.as_str()).collect();
+
+    // Join words, but don't add space before certain punctuation
+    let mut output = String::new();
+    for (i, word) in result.iter().enumerate() {
+        // Check if this word is standalone punctuation that shouldn't have space before it
+        // Contractions like "'re" or "'s" should have spaces before them
+        let is_standalone_punct = word.len() == 1 &&
+            word.chars().all(|c| matches!(c, '.' | ',' | '!' | '?' | ';' | ':' | ')'));
+
+        if i > 0 && !is_standalone_punct {
+            output.push(' ');
+        }
+        output.push_str(word);
+    }
+    output
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_word_grouping() {
+        let tokens = vec![
+            TimedToken {
+                text: "▁Hello".to_string(),
+                start: 0.0,
+                end: 0.5,
+            },
+            TimedToken {
+                text: "▁world".to_string(),
+                start: 0.5,
+                end: 1.0,
+            },
+        ];
+
+        let words = group_by_words(&tokens);
+        assert_eq!(words.len(), 2);
+        assert_eq!(words[0].text, "Hello");
+        assert_eq!(words[1].text, "world");
+    }
+
+    #[test]
+    fn test_sentence_grouping() {
+        let tokens = vec![
+            TimedToken {
+                text: "▁Hello".to_string(),
+                start: 0.0,
+                end: 0.5,
+            },
+            TimedToken {
+                text: "▁world".to_string(),
+                start: 0.5,
+                end: 1.0,
+            },
+            TimedToken {
+                text: ".".to_string(),
+                start: 1.0,
+                end: 1.1,
+            },
+        ];
+
+        let sentences = group_by_sentences(&tokens);
+        assert_eq!(sentences.len(), 1);
+        assert_eq!(sentences[0].text, "Hello world.");
+        assert_eq!(sentences[0].start, 0.0);
+        assert_eq!(sentences[0].end, 1.1);
+    }
+
+    #[test]
+    fn test_repetition_preservation() {
+        let words = vec![
+            TimedToken {
+                text: "uh".to_string(),
+                start: 0.0,
+                end: 0.5,
+            },
+            TimedToken {
+                text: "uh".to_string(),
+                start: 0.5,
+                end: 1.0,
+            },
+            TimedToken {
+                text: "hello".to_string(),
+                start: 1.0,
+                end: 1.5,
+            },
+        ];
+
+        let result = format_sentence(&words);
+        assert_eq!(result, "uh uh hello");
+    }
+}
diff --git a/vendor/parakeet-rs/src/vocab.rs b/vendor/parakeet-rs/src/vocab.rs
new file mode 100644
index 0000000..888568e
--- /dev/null
+++ b/vendor/parakeet-rs/src/vocab.rs
@@ -0,0 +1,63 @@
+use crate::error::{Error, Result};
+use std::fs::File;
+use std::io::{BufRead, BufReader};
+use std::path::Path;
+
+/// Vocabulary parser for vocab.txt format used by TDT models
+#[derive(Debug, Clone)]
+pub struct Vocabulary {
+    pub id_to_token: Vec<String>,
+    pub _blank_id: usize,
+}
+
+impl Vocabulary {
+    /// Load vocabulary from vocab.txt file
+    pub fn from_file<P: AsRef<Path>>(path: P) -> Result<Self> {
+        let file = File::open(path.as_ref()).map_err(|e| {
+            Error::Config(format!("Failed to open vocab file: {}", e))
+        })?;
+
+        let reader = BufReader::new(file);
+        let mut id_to_token = Vec::new();
+        let mut blank_id = 0;
+
+        for line in reader.lines() {
+            let line = line.map_err(|e| {
+                Error::Config(format!("Failed to read vocab file: {}", e))
+            })?;
+
+            let parts: Vec<&str> = line.splitn(2, ' ').collect();
+            if parts.len() == 2 {
+                let token = parts[0].to_string();
+                let id: usize = parts[1].parse().map_err(|e| {
+                    Error::Config(format!("Invalid token ID in vocab: {}", e))
+                })?;
+
+                if id >= id_to_token.len() {
+                    id_to_token.resize(id + 1, String::new());
+                }
+                id_to_token[id] = token.clone();
+
+                // Track blank token
+                if token == "<blk>" || token == "<blank>" {
+                    blank_id = id;
+                }
+            }
+        }
+
+        // Default to last token if no blank found
+        if blank_id == 0 && !id_to_token.is_empty() {
+            blank_id = id_to_token.len() - 1;
+        }
+
+        Ok(Self {
+            id_to_token,
+            _blank_id: blank_id,
+        })
+    }
+
+    /// Get token by ID
+    pub fn id_to_text(&self, id: usize) -> Option<&str> {
+        self.id_to_token.get(id).map(|s| s.as_str())
+    }
+}