Add Qwen3-TTS streaming endpoint for voice synthesis (#40)

* Task completion checkpoint

* Task completion checkpoint

* Task completion checkpoint

* Add Qwen3-TTS research document for live TTS replacement

Research findings for replacing Chatterbox TTS with Qwen3-TTS-12Hz-0.6B-Base:

- Current TTS: Chatterbox-Turbo-ONNX with batch-only generation, no streaming
- Qwen3-TTS: 97ms end-to-end latency, streaming support, 3-second voice cloning
- Voice cloning: Requires 3s reference audio + transcript (Makima voice planned)
- Integration: Python service with WebSocket bridge (no ONNX export available)
- Languages: 10 supported including English and Japanese

Document includes:
- Current architecture analysis (makima/src/tts.rs)
- Qwen3-TTS capabilities and requirements
- Feasibility assessment for live/streaming TTS
- Audio clip requirements for voice cloning
- Preliminary technical approach with architecture diagrams

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* [WIP] Heartbeat checkpoint - 2026-01-27 03:11:15 UTC

* Add Qwen3-TTS research documentation

Comprehensive research on replacing Chatterbox TTS with Qwen3-TTS-12Hz-0.6B-Base:

- Current TTS implementation analysis (Chatterbox-Turbo-ONNX in makima/src/tts.rs)
- Qwen3-TTS capabilities: 97ms streaming latency, voice cloning with 3s reference
- Cross-lingual support: Japanese voice (Makima/Tomori Kusunoki) speaking English
- Python microservice architecture recommendation (FastAPI + WebSocket)
- Implementation phases and technical approach
- Hardware requirements and dependencies

Key findings:
- Live/streaming TTS is highly feasible with 97ms latency
- Voice cloning fully supported with 0.95 speaker similarity
- Recommended: Python microservice with WebSocket streaming

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add comprehensive Qwen3-TTS integration specification

This specification document defines the complete integration of
Qwen3-TTS-12Hz-0.6B-Base as a replacement for the existing Chatterbox-Turbo
TTS implementation. The document covers:

## Functional Requirements
- WebSocket endpoint /api/v1/speak for streaming TTS
- Voice cloning with default Makima voice (Japanese VA speaking English)
- Support for custom voice references
- Detailed client-to-server and server-to-client message protocols
- Integration with Listen page for bidirectional speech

## Non-Functional Requirements
- Latency targets: < 200ms first audio byte
- Audio quality: 24kHz, mono, PCM16/PCM32f
- Hardware requirements: CUDA GPU with 4-8GB VRAM
- Scalability: 10 concurrent sessions per GPU

## Architecture Specification
- Python TTS microservice with FastAPI/WebSocket
- Rust proxy endpoint in makima server
- Voice prompt caching mechanism (LRU cache)
- Error handling and recovery strategies

## API Contract
- Complete WebSocket message format definitions (TypeScript)
- Error codes and responses (TTS_UNAVAILABLE, SYNTHESIS_ERROR, etc.)
- Session state machine and lifecycle management

## Voice Asset Requirements
- Makima voice clip specifications (5-10s WAV, transcript required)
- Storage location: models/voices/makima/
- Metadata format for voice management

## Testing Strategy
- Unit tests for Python TTS service and Rust proxy
- Integration tests for WebSocket flow
- Latency benchmarks with performance targets
- Test data fixtures for various text lengths

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add Qwen3-TTS implementation plan

Comprehensive implementation plan for replacing Chatterbox-TTS with
Qwen3-TTS streaming TTS service, including:

- Task breakdown with estimated hours for each phase
- Phase 1: Python TTS microservice (FastAPI, WebSocket)
- Phase 2: Rust proxy integration (speak.rs, tts_client.rs)
- Detailed file changes and new module structure
- Testing plan with unit, integration, and latency benchmarks
- Risk assessment with mitigation strategies
- Success criteria for each phase

Based on specification in docs/specs/qwen3-tts-spec.md

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add author and research references to TTS implementation plan

Add links to research documentation and author attribution.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* [WIP] Heartbeat checkpoint - 2026-01-27 03:25:06 UTC

* Add Python TTS service project structure (Phase 1.1-1.3)

Create the initial makima-tts Python service directory structure with:
- pyproject.toml with FastAPI, Qwen-TTS, and torch dependencies
- config.py with pydantic-settings TTSConfig class
- models.py with Pydantic message models (Start, Speak, Stop, Ready, etc.)

This implements tasks P1.1, P1.2, and P1.3 from the Qwen3-TTS implementation plan.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add TTS engine and voice manager for Qwen3-TTS (Phase 1.4-1.5)

Implement core TTS functionality:
- tts_engine.py: Qwen3-TTS wrapper with streaming audio chunk generation
- voice_manager.py: Voice prompt caching with LRU eviction and TTL support

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* [WIP] Heartbeat checkpoint - 2026-01-27 03:30:06 UTC

* Add TTS proxy client and message types (Phase 2.1, 2.2, 2.4)

- Add tts_client.rs with TtsConfig, TtsCircuitBreaker, TtsError,
  TtsProxyClient, and TtsConnection structs for WebSocket proxying
- Add TTS message types to messages.rs (TtsAudioEncoding, TtsPriority,
  TtsStartMessage, TtsSpeakMessage, TtsStopMessage, TtsClientMessage,
  TtsReadyMessage, TtsAudioChunkMessage, TtsCompleteMessage,
  TtsErrorMessage, TtsStoppedMessage, TtsServerMessage)
- Export tts_client module from server mod.rs
- tokio-tungstenite already present in Cargo.toml

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add TTS WebSocket handler and route (Phase 2.3, 2.5, 2.6)

- Create speak.rs WebSocket handler that proxies to Python TTS service
- Add TtsState fields (tts_client, tts_config) to AppState
- Add with_tts() builder and is_tts_healthy() methods to AppState
- Register /api/v1/speak route in the router
- Add speak module export in handlers/mod.rs

The handler forwards WebSocket messages bidirectionally between
the client and the Python TTS microservice with proper error handling.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add Makima voice profile assets for TTS voice cloning

Creates the voice assets directory structure with:
- manifest.json containing voice configuration (voice_id, speaker,
  language, reference audio path, and Japanese transcript placeholder)
- README.md with instructions for obtaining voice reference audio

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Add Rust-native Qwen3-TTS integration research document

Research findings for integrating Qwen3-TTS-12Hz-0.6B-Base directly into
the makima Rust codebase without Python. Key conclusions:
- ONNX export is not viable (unsupported architecture)
- Candle (HF Rust ML framework) is the recommended approach
- Model weights available in safetensors format (2.52GB total)
- Three components needed: LM backbone, code predictor, speech tokenizer
- Crane project has Qwen3-TTS as highest priority (potential upstream)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* [WIP] Heartbeat checkpoint - 2026-01-27 11:21:43 UTC

* [WIP] Heartbeat checkpoint - 2026-01-27 11:24:19 UTC

* [WIP] Heartbeat checkpoint - 2026-01-27 11:26:43 UTC

* feat: implement Rust-native Qwen3-TTS using candle framework

Replace monolithic tts.rs with modular tts/ directory structure:

- tts/mod.rs: TtsEngine trait, TtsEngineFactory, shared types (AudioChunk,
  TtsError), and utility functions (save_wav, resample, argmax)
- tts/chatterbox.rs: existing ONNX-based ChatterboxTTS adapted to implement
  TtsEngine trait with Mutex-wrapped sessions for Send+Sync
- tts/qwen3/mod.rs: Qwen3Tts entry point with HuggingFace model loading
- tts/qwen3/config.rs: Qwen3TtsConfig parsing from HF config.json
- tts/qwen3/model.rs: 28-layer Qwen3 transformer with RoPE, GQA (16 heads,
  8 KV heads), SiLU MLP, RMS norm, and KV cache
- tts/qwen3/code_predictor.rs: 5-layer MTP module predicting 16 codebooks
- tts/qwen3/speech_tokenizer.rs: ConvNet encoder/decoder with 16-layer RVQ
- tts/qwen3/generate.rs: autoregressive generation loop with streaming support

Add candle-core, candle-nn, candle-transformers, safetensors to Cargo.toml.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat: integrate TTS engine into speak WebSocket handler

- Update speak.rs handler to use TTS engine directly from SharedState
  instead of returning a stub "not implemented" error
- Add TtsEngine (OnceCell lazy-loaded) to AppState in state.rs with
  get_tts_engine() method for lazy initialization on first connection
- Implement full WebSocket protocol: client sends JSON speak/cancel/stop
  messages, server streams binary PCM audio chunks and audio_end signals
- Create voices/makima/manifest.json for Makima voice profile configuration
- All files compile successfully with zero errors

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* feat: add /speak TTS page with WebSocket audio playback

Add a new /speak frontend page for text-to-speech via WebSocket.
The page accepts text input and streams synthesized PCM audio through
the Web Audio API. Includes model loading indicator, cancel support,
and connection status. Also adds a loading bar to the listen page
ControlPanel during WebSocket connection.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

1 files changed, 485 insertions, 0 deletions

diff --git a/makima/src/tts/chatterbox.rs b/makima/src/tts/chatterbox.rs
new file mode 100644
index 0000000..e26bc06
--- /dev/null
+++ b/makima/src/tts/chatterbox.rs
@@ -0,0 +1,485 @@
+//! Chatterbox TTS engine — ONNX-based (legacy).
+//!
+//! This is the existing Chatterbox TTS implementation moved from `tts.rs`,
+//! now implementing the `TtsEngine` trait for unified access.
+
+use std::borrow::Cow;
+use std::fs;
+use std::path::{Path, PathBuf};
+use std::sync::Mutex;
+
+use hf_hub::api::sync::Api;
+use ndarray::{Array2, Array3, Array4, ArrayD, IxDyn};
+use ort::session::Session;
+use ort::value::{DynValue, Value};
+use tokenizers::Tokenizer;
+
+use crate::audio;
+
+use super::{
+    apply_repetition_penalty, argmax, resample_to_24k, AudioChunk, TtsEngine, TtsError,
+    SAMPLE_RATE,
+};
+
+const START_SPEECH_TOKEN: i64 = 6561;
+const STOP_SPEECH_TOKEN: i64 = 6562;
+const SILENCE_TOKEN: i64 = 4299;
+const NUM_LAYERS: usize = 24;
+const NUM_KV_HEADS: usize = 16;
+const HEAD_DIM: usize = 64;
+
+const MODEL_ID: &str = "ResembleAI/chatterbox-turbo-ONNX";
+const DEFAULT_MODEL_DIR: &str = "models/chatterbox-turbo";
+
+struct VoiceCondition {
+    audio_features: ArrayD<f32>,
+    prompt_tokens: ArrayD<i64>,
+    speaker_embeddings: ArrayD<f32>,
+    speaker_features: ArrayD<f32>,
+}
+
+fn extract_f32_tensor(value: &Value) -> Result<ArrayD<f32>, TtsError> {
+    let (shape, data) = value
+        .try_extract_tensor::<f32>()
+        .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+    let dims: Vec<usize> = shape.iter().map(|&d| d as usize).collect();
+    ArrayD::from_shape_vec(IxDyn(&dims), data.to_vec())
+        .map_err(|e| TtsError::Inference(e.to_string()))
+}
+
+fn extract_i64_tensor(value: &Value) -> Result<ArrayD<i64>, TtsError> {
+    let (shape, data) = value
+        .try_extract_tensor::<i64>()
+        .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+    let dims: Vec<usize> = shape.iter().map(|&d| d as usize).collect();
+    ArrayD::from_shape_vec(IxDyn(&dims), data.to_vec())
+        .map_err(|e| TtsError::Inference(e.to_string()))
+}
+
+pub struct ChatterboxTTS {
+    speech_encoder: Mutex<Session>,
+    embed_tokens: Mutex<Session>,
+    language_model: Mutex<Session>,
+    conditional_decoder: Mutex<Session>,
+    tokenizer: Tokenizer,
+}
+
+// SAFETY: Sessions are behind Mutex, Tokenizer is Send+Sync
+unsafe impl Send for ChatterboxTTS {}
+unsafe impl Sync for ChatterboxTTS {}
+
+impl ChatterboxTTS {
+    pub fn from_pretrained(model_dir: Option<&str>) -> Result<Self, TtsError> {
+        let model_path = PathBuf::from(model_dir.unwrap_or(DEFAULT_MODEL_DIR));
+
+        if !model_path.exists() {
+            download_models(&model_path)?;
+        }
+
+        Self::load_from_path(&model_path)
+    }
+
+    pub fn load_from_path(model_dir: &Path) -> Result<Self, TtsError> {
+        let speech_encoder = Session::builder()?
+            .with_intra_threads(4)?
+            .commit_from_file(model_dir.join("speech_encoder.onnx"))?;
+
+        let embed_tokens = Session::builder()?
+            .with_intra_threads(4)?
+            .commit_from_file(model_dir.join("embed_tokens.onnx"))?;
+
+        let language_model = Session::builder()?
+            .with_intra_threads(4)?
+            .commit_from_file(model_dir.join("language_model.onnx"))?;
+
+        let conditional_decoder = Session::builder()?
+            .with_intra_threads(4)?
+            .commit_from_file(model_dir.join("conditional_decoder.onnx"))?;
+
+        let tokenizer_path = model_dir.join("tokenizer.json");
+        let tokenizer = Tokenizer::from_file(&tokenizer_path)
+            .map_err(|e| TtsError::Tokenizer(e.to_string()))?;
+
+        Ok(Self {
+            speech_encoder: Mutex::new(speech_encoder),
+            embed_tokens: Mutex::new(embed_tokens),
+            language_model: Mutex::new(language_model),
+            conditional_decoder: Mutex::new(conditional_decoder),
+            tokenizer,
+        })
+    }
+
+    pub fn generate_tts(&self) -> Result<Vec<f32>, TtsError> {
+        Err(TtsError::VoiceRequired)
+    }
+
+    pub fn generate_tts_with_voice(
+        &self,
+        text: &str,
+        sample_audio_path: &Path,
+    ) -> Result<Vec<f32>, TtsError> {
+        let audio = audio::to_16k_mono_from_path(sample_audio_path)?;
+        let resampled = resample_to_24k(&audio.samples, audio.sample_rate);
+        self.generate_tts_with_samples(text, &resampled, SAMPLE_RATE)
+    }
+
+    pub fn generate_tts_with_samples(
+        &self,
+        text: &str,
+        samples: &[f32],
+        sample_rate: u32,
+    ) -> Result<Vec<f32>, TtsError> {
+        let resampled = if sample_rate != SAMPLE_RATE {
+            resample_to_24k(samples, sample_rate)
+        } else {
+            samples.to_vec()
+        };
+
+        let voice_condition = self.encode_voice(&resampled)?;
+
+        let encoding = self
+            .tokenizer
+            .encode(text, true)
+            .map_err(|e| TtsError::Tokenizer(e.to_string()))?;
+
+        let text_input_ids: Vec<i64> = encoding.get_ids().iter().map(|&id| id as i64).collect();
+
+        let generated_tokens =
+            self.generate_speech_tokens(&text_input_ids, &voice_condition.audio_features)?;
+
+        let prompt_tokens: Vec<i64> = voice_condition.prompt_tokens.iter().copied().collect();
+        let silence_tokens = vec![SILENCE_TOKEN; 3];
+
+        let mut final_tokens =
+            Vec::with_capacity(prompt_tokens.len() + generated_tokens.len() + silence_tokens.len());
+        final_tokens.extend_from_slice(&prompt_tokens);
+        final_tokens.extend_from_slice(&generated_tokens);
+        final_tokens.extend_from_slice(&silence_tokens);
+
+        let audio_samples = self.decode_speech_tokens(
+            &final_tokens,
+            &voice_condition.speaker_embeddings,
+            &voice_condition.speaker_features,
+        )?;
+
+        Ok(audio_samples)
+    }
+
+    fn encode_voice(&self, samples: &[f32]) -> Result<VoiceCondition, TtsError> {
+        let audio_arr = Array2::from_shape_vec((1, samples.len()), samples.to_vec())
+            .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+        let audio_tensor = Value::from_array(audio_arr)?;
+
+        let mut encoder = self
+            .speech_encoder
+            .lock()
+            .map_err(|e| TtsError::Inference(e.to_string()))?;
+        let outputs = encoder.run(ort::inputs!["audio_values" => audio_tensor])?;
+
+        let audio_features = extract_f32_tensor(&outputs[0])?;
+        let prompt_tokens = extract_i64_tensor(&outputs[1])?;
+        let speaker_embeddings = extract_f32_tensor(&outputs[2])?;
+        let speaker_features = extract_f32_tensor(&outputs[3])?;
+
+        Ok(VoiceCondition {
+            audio_features,
+            prompt_tokens,
+            speaker_embeddings,
+            speaker_features,
+        })
+    }
+
+    fn generate_speech_tokens(
+        &self,
+        text_input_ids: &[i64],
+        audio_features: &ArrayD<f32>,
+    ) -> Result<Vec<i64>, TtsError> {
+        let max_new_tokens: usize = 1024;
+        let repetition_penalty: f32 = 1.2;
+
+        let mut generate_tokens: Vec<i64> = vec![START_SPEECH_TOKEN];
+
+        let mut past_key_values = Self::init_kv_cache(0);
+        let mut first_iteration = true;
+        let mut total_seq_len: usize = 0;
+
+        for _ in 0..max_new_tokens {
+            let current_input_ids = if first_iteration {
+                text_input_ids.to_vec()
+            } else {
+                vec![*generate_tokens.last().unwrap()]
+            };
+
+            let input_ids_arr =
+                Array2::from_shape_vec((1, current_input_ids.len()), current_input_ids)
+                    .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+            let input_ids_tensor = Value::from_array(input_ids_arr)?;
+
+            let inputs_embeds = {
+                let mut embed = self
+                    .embed_tokens
+                    .lock()
+                    .map_err(|e| TtsError::Inference(e.to_string()))?;
+                let embed_outputs = embed.run(ort::inputs![input_ids_tensor])?;
+                extract_f32_tensor(&embed_outputs[0])?
+            };
+
+            let inputs_embeds = if first_iteration {
+                let audio_feat_3d = audio_features
+                    .view()
+                    .into_dimensionality::<ndarray::Ix3>()
+                    .map_err(|e| TtsError::Inference(e.to_string()))?;
+                let text_emb_3d = inputs_embeds
+                    .view()
+                    .into_dimensionality::<ndarray::Ix3>()
+                    .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+                ndarray::concatenate(ndarray::Axis(1), &[audio_feat_3d, text_emb_3d])
+                    .map_err(|e| TtsError::Inference(e.to_string()))?
+            } else {
+                inputs_embeds
+                    .view()
+                    .into_dimensionality::<ndarray::Ix3>()
+                    .map_err(|e| TtsError::Inference(e.to_string()))?
+                    .to_owned()
+            };
+
+            let seq_len = inputs_embeds.shape()[1];
+
+            let (attention_mask, position_ids) = if first_iteration {
+                total_seq_len = seq_len;
+                let attention_mask: Array2<i64> = Array2::ones((1, seq_len));
+                let position_ids = Array2::from_shape_fn((1, seq_len), |(_, j)| j as i64);
+                (attention_mask, position_ids)
+            } else {
+                total_seq_len += 1;
+                let attention_mask: Array2<i64> = Array2::ones((1, total_seq_len));
+                let position_ids =
+                    Array2::from_shape_vec((1, 1), vec![(total_seq_len - 1) as i64])
+                        .map_err(|e| TtsError::Inference(e.to_string()))?;
+                (attention_mask, position_ids)
+            };
+
+            let (logits, new_kv) = self.run_language_model(
+                inputs_embeds,
+                position_ids,
+                attention_mask,
+                past_key_values,
+            )?;
+
+            past_key_values = new_kv;
+
+            let logits_3d = logits
+                .view()
+                .into_dimensionality::<ndarray::Ix3>()
+                .map_err(|e| TtsError::Inference(e.to_string()))?;
+            let last_idx = logits_3d.shape()[1] - 1;
+
+            let mut current_logits: Vec<f32> = logits_3d
+                .slice(ndarray::s![0, last_idx, ..])
+                .iter()
+                .copied()
+                .collect();
+
+            apply_repetition_penalty(&mut current_logits, &generate_tokens, repetition_penalty);
+
+            let next_token = argmax(&current_logits);
+            generate_tokens.push(next_token);
+
+            if next_token == STOP_SPEECH_TOKEN {
+                break;
+            }
+
+            first_iteration = false;
+        }
+
+        if generate_tokens.len() > 2 {
+            Ok(generate_tokens[1..generate_tokens.len() - 1].to_vec())
+        } else {
+            Ok(Vec::new())
+        }
+    }
+
+    fn init_kv_cache(seq_len: usize) -> Vec<Array4<f32>> {
+        let mut cache = Vec::with_capacity(NUM_LAYERS * 2);
+        for _ in 0..NUM_LAYERS {
+            let key = Array4::<f32>::zeros((1, NUM_KV_HEADS, seq_len, HEAD_DIM));
+            let value = Array4::<f32>::zeros((1, NUM_KV_HEADS, seq_len, HEAD_DIM));
+            cache.push(key);
+            cache.push(value);
+        }
+        cache
+    }
+
+    fn run_language_model(
+        &self,
+        inputs_embeds: Array3<f32>,
+        position_ids: Array2<i64>,
+        attention_mask: Array2<i64>,
+        past_key_values: Vec<Array4<f32>>,
+    ) -> Result<(ArrayD<f32>, Vec<Array4<f32>>), TtsError> {
+        let mut inputs: Vec<(Cow<str>, DynValue)> = Vec::new();
+
+        inputs.push((
+            Cow::from("inputs_embeds"),
+            Value::from_array(inputs_embeds)?.into_dyn(),
+        ));
+        inputs.push((
+            Cow::from("position_ids"),
+            Value::from_array(position_ids)?.into_dyn(),
+        ));
+        inputs.push((
+            Cow::from("attention_mask"),
+            Value::from_array(attention_mask)?.into_dyn(),
+        ));
+
+        for layer_idx in 0..NUM_LAYERS {
+            let key_name = format!("past_key_values.{}.key", layer_idx);
+            let value_name = format!("past_key_values.{}.value", layer_idx);
+
+            let key_tensor =
+                Value::from_array(past_key_values[layer_idx * 2].clone())?.into_dyn();
+            let value_tensor =
+                Value::from_array(past_key_values[layer_idx * 2 + 1].clone())?.into_dyn();
+
+            inputs.push((Cow::from(key_name), key_tensor));
+            inputs.push((Cow::from(value_name), value_tensor));
+        }
+
+        let mut lm = self
+            .language_model
+            .lock()
+            .map_err(|e| TtsError::Inference(e.to_string()))?;
+        let outputs = lm.run(inputs)?;
+
+        let logits = extract_f32_tensor(&outputs[0])?;
+
+        let mut new_kv = Vec::with_capacity(NUM_LAYERS * 2);
+        for layer_idx in 0..NUM_LAYERS {
+            let key_idx = 1 + layer_idx * 2;
+            let value_idx = 2 + layer_idx * 2;
+
+            let key_arr = extract_f32_tensor(&outputs[key_idx])?;
+            let value_arr = extract_f32_tensor(&outputs[value_idx])?;
+
+            let key_4d = key_arr
+                .into_dimensionality::<ndarray::Ix4>()
+                .map_err(|e| TtsError::Inference(e.to_string()))?;
+            let value_4d = value_arr
+                .into_dimensionality::<ndarray::Ix4>()
+                .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+            new_kv.push(key_4d.to_owned());
+            new_kv.push(value_4d.to_owned());
+        }
+
+        Ok((logits, new_kv))
+    }
+
+    fn decode_speech_tokens(
+        &self,
+        speech_tokens: &[i64],
+        speaker_embeddings: &ArrayD<f32>,
+        speaker_features: &ArrayD<f32>,
+    ) -> Result<Vec<f32>, TtsError> {
+        if speech_tokens.is_empty() {
+            return Ok(Vec::new());
+        }
+
+        let tokens_arr =
+            Array2::from_shape_vec((1, speech_tokens.len()), speech_tokens.to_vec())
+                .map_err(|e| TtsError::Inference(e.to_string()))?;
+
+        let mut inputs: Vec<(Cow<str>, DynValue)> = Vec::new();
+        inputs.push((
+            Cow::from("speech_tokens"),
+            Value::from_array(tokens_arr)?.into_dyn(),
+        ));
+        inputs.push((
+            Cow::from("speaker_embeddings"),
+            Value::from_array(speaker_embeddings.clone())?.into_dyn(),
+        ));
+        inputs.push((
+            Cow::from("speaker_features"),
+            Value::from_array(speaker_features.clone())?.into_dyn(),
+        ));
+
+        let mut decoder = self
+            .conditional_decoder
+            .lock()
+            .map_err(|e| TtsError::Inference(e.to_string()))?;
+        let outputs = decoder.run(inputs)?;
+
+        let waveform = extract_f32_tensor(&outputs[0])?;
+
+        Ok(waveform.iter().copied().collect())
+    }
+}
+
+#[async_trait::async_trait]
+impl TtsEngine for ChatterboxTTS {
+    async fn generate(
+        &self,
+        text: &str,
+        reference_audio: Option<&[f32]>,
+        reference_sample_rate: Option<u32>,
+    ) -> Result<Vec<AudioChunk>, TtsError> {
+        let samples = match reference_audio {
+            Some(audio) => {
+                let sr = reference_sample_rate.unwrap_or(SAMPLE_RATE);
+                self.generate_tts_with_samples(text, audio, sr)?
+            }
+            None => return Err(TtsError::VoiceRequired),
+        };
+
+        Ok(vec![AudioChunk {
+            samples,
+            sample_rate: SAMPLE_RATE,
+            is_final: true,
+        }])
+    }
+
+    fn is_ready(&self) -> bool {
+        true
+    }
+}
+
+fn download_models(target_dir: &Path) -> Result<(), TtsError> {
+    fs::create_dir_all(target_dir)?;
+
+    let api = Api::new().map_err(|e| TtsError::ModelLoad(e.to_string()))?;
+    let repo = api.model(MODEL_ID.to_string());
+
+    let model_files = [
+        "onnx/speech_encoder.onnx",
+        "onnx/speech_encoder.onnx_data",
+        "onnx/embed_tokens.onnx",
+        "onnx/embed_tokens.onnx_data",
+        "onnx/language_model.onnx",
+        "onnx/language_model.onnx_data",
+        "onnx/conditional_decoder.onnx",
+        "onnx/conditional_decoder.onnx_data",
+        "tokenizer.json",
+    ];
+
+    for file in &model_files {
+        println!("Downloading {}...", file);
+        let downloaded_path = repo
+            .get(file)
+            .map_err(|e| TtsError::ModelLoad(e.to_string()))?;
+
+        let filename = Path::new(file).file_name().unwrap();
+        let target_path = target_dir.join(filename);
+
+        if !target_path.exists() {
+            fs::copy(&downloaded_path, &target_path)?;
+        }
+    }
+
+    println!("Models downloaded to {:?}", target_dir);
+    Ok(())
+}