fix: address self-review and CodeRabbit findings

Critical fixes:
- Add tts_backend parameter to create_voice_prompt_for_profile (was crashing Hebrew generation)
- Fix double-wrapped language tokens in Whisper suppress_tokens (was producing malformed IDs)
- Add threading lock to torch.load monkey-patch for thread safety

High priority:
- Deduplicate STT_MODEL_MAP into backends/__init__.py (was copy-pasted in mlx + pytorch)
- Scope trim_tts_output to Chatterbox only (was aggressively trimming Qwen output)
- Expand TranscriptionRequest language pattern to all 11 supported languages
- Add Chatterbox sub-dependencies to requirements.txt (conformer, diffusers, etc.)
- Read sample_rate from Chatterbox model object instead of hardcoding 24000

Cleanup:
- Remove duplicate import asyncio
- Remove console.log debug statements from client.ts triggerModelDownload
- Add CUDA cache cleanup to Chatterbox unload_model
- Add chatterbox unload to shutdown handler
- Fix unused variable warnings (sr -> _sr)
- Fix f-strings without placeholders

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: nadavox

app/src/lib/api/client.ts

@@ -310,13 +310,10 @@ class ApiClient {
   }
 
   async triggerModelDownload(modelName: string): Promise<{ message: string }> {
-    console.log('[API] triggerModelDownload called for:', modelName, 'at', new Date().toISOString());
-    const result = await this.request<{ message: string }>('/models/download', {
+    return this.request<{ message: string }>('/models/download', {
       method: 'POST',
       body: JSON.stringify({ model_name: modelName } as ModelDownloadRequest),
     });
-    console.log('[API] triggerModelDownload response:', result);
-    return result;
   }
 
   async deleteModel(modelName: string): Promise<{ message: string }> {

backend/backends/__init__.py

@@ -10,6 +10,17 @@
 
 from ..platform_detect import get_backend_type
 
+# Shared model name mapping for STT backends (MLX + PyTorch).
+# Maps short model size keys to HuggingFace repo IDs.
+STT_MODEL_MAP = {
+    "base": "openai/whisper-base",
+    "small": "openai/whisper-small",
+    "medium": "openai/whisper-medium",
+    "large": "openai/whisper-large",
+    "ivrit-v3": "ivrit-ai/whisper-large-v3",
+    "ivrit-v3-turbo": "ivrit-ai/whisper-large-v3-turbo",
+}
+
 
 @runtime_checkable
 class TTSBackend(Protocol):

backend/backends/chatterbox_backend.py

@@ -136,17 +136,20 @@ def _load_model_sync(self, model_size: str):
                 # The multilingual model's .pt files were saved on CUDA and
                 # from_local() doesn't pass map_location, so loading on CPU fails.
                 if device == "cpu":
+                    import threading
                     _orig_torch_load = torch.load
+                    _load_lock = threading.Lock()
 
                     def _patched_load(*args, **kwargs):
                         kwargs.setdefault("map_location", "cpu")
                         return _orig_torch_load(*args, **kwargs)
 
-                    torch.load = _patched_load
-                    try:
-                        self.model = ChatterboxMultilingualTTS.from_pretrained(device=device)
-                    finally:
-                        torch.load = _orig_torch_load
+                    with _load_lock:
+                        torch.load = _patched_load
+                        try:
+                            self.model = ChatterboxMultilingualTTS.from_pretrained(device=device)
+                        finally:
+                            torch.load = _orig_torch_load
                 else:
                     self.model = ChatterboxMultilingualTTS.from_pretrained(device=device)
 
@@ -171,8 +174,8 @@ def _patched_load(*args, **kwargs):
 
         except ImportError as e:
             print(
-                f"Error: chatterbox-tts package not found. "
-                f"Install with: pip install chatterbox-tts"
+                "Error: chatterbox-tts package not found. "
+                "Install with: pip install chatterbox-tts"
             )
             progress_manager = get_progress_manager()
             task_manager = get_task_manager()
@@ -218,9 +221,13 @@ def _patched_add_hebrew_diacritics(text: str) -> str:
     def unload_model(self) -> None:
         """Unload model to free memory."""
         if self.model is not None:
+            device = self._device
             del self.model
             self.model = None
             self._device = None
+            if device == "cuda":
+                import torch
+                torch.cuda.empty_cache()
             print("Chatterbox Multilingual TTS model unloaded")
 
     async def create_voice_prompt(
@@ -250,7 +257,7 @@ async def combine_voice_prompts(
         combined_audio = []
 
         for audio_path in audio_paths:
-            audio, sr = load_audio(audio_path)
+            audio, _sr = load_audio(audio_path)
             audio = normalize_audio(audio)
             combined_audio.append(audio)
 
@@ -334,8 +341,7 @@ def _generate_sync():
             else:
                 audio = np.asarray(wav, dtype=np.float32)
 
-            # Chatterbox default sample rate is 24000
-            sample_rate = 24000
+            sample_rate = getattr(self.model, 'sr', None) or getattr(self.model, 'sample_rate', 24000)
 
             return audio, sample_rate
 

backend/backends/mlx_backend.py

@@ -400,14 +400,7 @@ async def generate_with_adapter(
         return await self.generate(text, voice_prompt, language, seed, instruct)
 
 
-STT_MODEL_MAP = {
-    "base": "openai/whisper-base",
-    "small": "openai/whisper-small",
-    "medium": "openai/whisper-medium",
-    "large": "openai/whisper-large",
-    "ivrit-v3": "ivrit-ai/whisper-large-v3",
-    "ivrit-v3-turbo": "ivrit-ai/whisper-large-v3-turbo",
-}
+from . import STT_MODEL_MAP
 
 
 class MLXSTTBackend:

backend/backends/pytorch_backend.py

@@ -369,14 +369,7 @@ def _generate_sync():
         return audio, sample_rate
 
 
-STT_MODEL_MAP = {
-    "base": "openai/whisper-base",
-    "small": "openai/whisper-small",
-    "medium": "openai/whisper-medium",
-    "large": "openai/whisper-large",
-    "ivrit-v3": "ivrit-ai/whisper-large-v3",
-    "ivrit-v3-turbo": "ivrit-ai/whisper-large-v3-turbo",
-}
+from . import STT_MODEL_MAP
 
 
 class PyTorchSTTBackend:
@@ -608,15 +601,13 @@ def _transcribe_sync():
                 tokenizer = self.processor.tokenizer
                 lang_token = f"<|{language}|>"
                 if lang_token in tokenizer.get_vocab():
-                    lang_id = tokenizer.convert_tokens_to_ids(lang_token)
                     # Suppress all other language tokens to prevent drift
                     all_lang_tokens = [
-                        tokenizer.convert_tokens_to_ids(f"<|{lang}|>")
-                        for lang in tokenizer.additional_special_tokens
-                        if lang.startswith("<|") and lang.endswith("|>")
-                        and lang != lang_token and lang != "<|transcribe|>"
-                        and lang != "<|notimestamps|>"
-                        and tokenizer.convert_tokens_to_ids(lang) != tokenizer.unk_token_id
+                        tokenizer.convert_tokens_to_ids(tok)
+                        for tok in tokenizer.additional_special_tokens
+                        if tok.startswith("<|") and tok.endswith("|>")
+                        and tok not in (lang_token, "<|transcribe|>", "<|notimestamps|>")
+                        and tokenizer.convert_tokens_to_ids(tok) != tokenizer.unk_token_id
                     ]
                     if all_lang_tokens:
                         generate_kwargs["suppress_tokens"] = all_lang_tokens

backend/main.py

@@ -19,7 +19,6 @@
 import io
 from pathlib import Path
 import uuid
-import asyncio
 import signal
 import os
 from urllib.parse import quote
@@ -704,9 +703,10 @@ async def download_model_background():
                 data.instruct,
             )
 
-        # Trim trailing silence/noise from TTS output (known Chatterbox issue)
-        from .utils.audio import trim_tts_output
-        audio = trim_tts_output(audio, sample_rate)
+        # Trim trailing silence/noise from Chatterbox output (known hallucination issue)
+        if data.language == "he":
+            from .utils.audio import trim_tts_output
+            audio = trim_tts_output(audio, sample_rate)
 
         # Calculate duration
         duration = len(audio) / sample_rate
@@ -2010,6 +2010,7 @@ async def shutdown_event():
     print("voicebox API shutting down...")
     # Unload models to free memory
     tts.unload_tts_model()
+    tts.unload_chatterbox_model()
     transcribe.unload_whisper_model()
 
 

backend/models.py

@@ -108,7 +108,7 @@ class HistoryListResponse(BaseModel):
 
 class TranscriptionRequest(BaseModel):
     """Request model for audio transcription."""
-    language: Optional[str] = Field(None, pattern="^(en|zh|he)$")
+    language: Optional[str] = Field(None, pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he)$")
 
 
 class TranscriptionResponse(BaseModel):

backend/profiles.py

@@ -327,6 +327,7 @@ async def create_voice_prompt_for_profile(
     profile_id: str,
     db: Session,
     use_cache: bool = True,
+    tts_backend=None,
 ) -> dict:
     """
     Create a combined voice prompt from all samples in a profile.
@@ -335,6 +336,7 @@ async def create_voice_prompt_for_profile(
         profile_id: Profile ID
         db: Database session
         use_cache: Whether to use cached prompts
+        tts_backend: Optional TTS backend override (e.g. Chatterbox for Hebrew)
 
     Returns:
         Voice prompt dictionary
@@ -345,12 +347,12 @@ async def create_voice_prompt_for_profile(
     if not samples:
         raise ValueError(f"No samples found for profile {profile_id}")
 
-    tts_model = get_tts_model()
+    backend = tts_backend or get_tts_model()
 
     if len(samples) == 1:
         # Single sample - use directly
         sample = samples[0]
-        voice_prompt, _ = await tts_model.create_voice_prompt(
+        voice_prompt, _ = await backend.create_voice_prompt(
             sample.audio_path,
             sample.reference_text,
             use_cache=use_cache,
@@ -362,7 +364,7 @@ async def create_voice_prompt_for_profile(
         reference_texts = [s.reference_text for s in samples]
 
         # Combine audio
-        combined_audio, combined_text = await tts_model.combine_voice_prompts(
+        combined_audio, combined_text = await backend.combine_voice_prompts(
             audio_paths,
             reference_texts,
         )

backend/requirements.txt

@@ -13,10 +13,16 @@ transformers>=4.36.0
 accelerate>=0.26.0
 huggingface_hub>=0.20.0
 qwen-tts>=0.0.5
-# Hebrew TTS — install with: pip install chatterbox-tts --no-deps
-# then install missing sub-deps: pip install conformer diffusers omegaconf pykakasi resemble-perth s3tokenizer
-# (numpy constraint on PyPI is too strict; works fine with numpy 2.x in practice)
+# Hebrew TTS (Chatterbox)
+# Note: chatterbox-tts has a strict numpy<2 pin on PyPI but works fine with numpy 2.x.
+# If pip fails, install with: pip install chatterbox-tts --no-deps
 chatterbox-tts>=0.1.0
+conformer
+diffusers
+omegaconf
+pykakasi
+resemble-perth
+s3tokenizer
 
 # Audio processing
 librosa>=0.10.0

backend/utils/audio.py

@@ -82,7 +82,7 @@ def save_audio(
 
 def prepare_for_transcription(
     audio: np.ndarray,
-    sr: int,
+    sr: int = 16000,  # noqa: ARG001 — kept for API consistency
 ) -> np.ndarray:
     """
     Prepare audio for Whisper transcription.