refactor: consolidate to Base 1.7B only, remove CustomVoice

- Remove get_custom_voice_model() and _MODEL_MODE checks
- Add profile_id_override param to create_voice_profile()
- Update Profile page: 2 models, Base 1.7B as single TTS engine
- Upload vivian_reference.wav to HF Space

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

app.py

@@ -409,9 +409,9 @@ else:
     print("[MomsVoice] No saved voice profile — using stock voice as default.")
 
 # Preload models at startup for fast response
-print("[MomsVoice] Preloading Qwen3-TTS 0.6B CustomVoice model...")
-from voice_clone import get_custom_voice_model
-get_custom_voice_model()
+print("[MomsVoice] Preloading Qwen3-TTS 1.7B Base model (cloning + stock voice)...")
+from voice_clone import get_qwen_tts
+get_qwen_tts()
 print("[MomsVoice] TTS model ready.")
 
 print("[MomsVoice] Preloading LFM2.5-Audio-1.5B Q&A model...")
@@ -419,38 +419,6 @@ from inference_lfm import get_lfm_model
 get_lfm_model()
 print("[MomsVoice] LFM Q&A model ready. All models preloaded.")
 
-# Pre-generate first paragraph audio for all stories (stock voice) for instant start
-import threading
-
-def _warmup_first_paragraphs():
-    """Background: pre-generate first 5 paragraphs of each story with stock voice to disk cache."""
-    import logging as _log
-    from tts import _get_cached_audio, _save_cached_audio, _synthesize_single
-    _logger = _log.getLogger("warmup")
-    WARMUP_PARAS = 5
-    for book in mock_books:
-        try:
-            paras = load_paragraphs(book["story_path"])
-            if not paras:
-                continue
-            # Pre-generate first 5 paragraphs
-            target_paras = paras[:WARMUP_PARAS]
-            all_chunks = split_into_chunks("\n\n".join(target_paras))
-            generated = 0
-            for chunk in all_chunks:
-                if _get_cached_audio(chunk, None) is not None:
-                    continue  # Already cached
-                wav, sr = _synthesize_single(chunk, None)
-                _save_cached_audio(chunk, None, wav, sr)
-                generated += 1
-            _logger.info("Warmed: %s (%d chunks, %d new)", book["title"], len(all_chunks), generated)
-        except Exception as e:
-            _logger.warning("Warmup failed for %s: %s", book.get("title", "?"), e)
-    _logger.info("All stories warmed (first %d paragraphs each).", WARMUP_PARAS)
-
-threading.Thread(target=_warmup_first_paragraphs, daemon=True).start()
-print("[MomsVoice] Background warmup started: pre-generating first 5 paragraphs for all stories.")
-
 
 # Gradio Application Core setup
 with gr.Blocks(title="MomsVoice", css=css_code) as demo:
@@ -661,12 +629,12 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
                             </div>
                             <div style="background: white; border: 1px solid #ebdccb; border-radius: 12px; padding: 10px;">
                                 <span style="font-size: 18px; display: block;">🧠</span>
-                                <strong style="font-size: 15px; display: block; color: #1c1c19;">3</strong>
+                                <strong style="font-size: 15px; display: block; color: #1c1c19;">2</strong>
                                 <span style="font-size: 9px; text-transform: uppercase; color: #6f6257;">Models</span>
                             </div>
                             <div style="background: white; border: 1px solid #ebdccb; border-radius: 12px; padding: 10px;">
                                 <span style="font-size: 18px; display: block;">🎙️</span>
-                                <strong style="font-size: 15px; display: block; color: #1c1c19;">Qwen3</strong>
+                                <strong style="font-size: 15px; display: block; color: #1c1c19;">Base 1.7B</strong>
                                 <span style="font-size: 9px; text-transform: uppercase; color: #6f6257;">TTS Engine</span>
                             </div>
                         </div>
@@ -681,8 +649,8 @@ with gr.Blocks(title="MomsVoice", css=css_code) as demo:
                     gr.HTML("""
                         <div style="padding: 12px; background: white; border-radius: 12px; border: 1px solid #ebdccb;">
                             <div style="font-size: 11.5px; color: #1c1c19; line-height: 2;">
-                                🎙️ <strong>Voice Cloning:</strong> Qwen3-TTS-1.7B-Base (bfloat16)<br>
-                                🗣️ <strong>Stock TTS:</strong> Qwen3-TTS-0.6B-CustomVoice<br>
+                                🎙️ <strong>TTS + Cloning:</strong> Qwen3-TTS-1.7B-Base (single model)<br>
+                                🗣️ <strong>Stock Voice:</strong> Vivian (via reference audio)<br>
                                 🧠 <strong>Q&A (end-to-end):</strong> LFM2.5-Audio-1.5B (LiquidAI)
                             </div>
                         </div>

test_qa_audio_check.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0f03ec6d9134fcdf7114ed7d1795d35612ff20b4618d3701f95443d15fdfe7e
+size 180524

voice_clone.py

@@ -46,13 +46,14 @@ torch.set_float32_matmul_precision("high")
 # ---------------------------------------------------------------------------
 # Model configuration
 # ---------------------------------------------------------------------------
-# Base model for zero-shot voice cloning (1.7B)
+# Base model for zero-shot voice cloning (1.7B) — used for BOTH cloned and stock voice
 BASE_MODEL_ID = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
-# CustomVoice model for fast predefined speakers (0.6B, ~3x faster)
-CUSTOM_VOICE_MODEL_ID = "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice"
 
-# Choose model via env var: "base" (default) or "custom_voice"
-_MODEL_MODE = os.environ.get("QWEN_TTS_MODE", "base").lower()
+# Stock voice reference audio (pre-generated "vivian" sample)
+VIVIAN_REF_PATH = Path(__file__).parent / "assets" / "vivian_reference.wav"
+
+# Profile ID for the built-in stock voice
+STOCK_VOICE_PROFILE_ID = "__stock_vivian__"
 
 # Optimized generation parameters (reduced from defaults: top_k=50, temp=0.9, max=2048)
 GENERATION_PARAMS = dict(
@@ -83,9 +84,7 @@ _PROFILE_CACHE: dict[str, list] = {}
 _cache_lock = threading.Lock()
 
 _qwen_tts_model = None
-_custom_voice_model = None
 _model_lock = threading.Lock()
-_custom_model_lock = threading.Lock()
 
 
 def _select_dtype() -> torch.dtype:
@@ -131,27 +130,6 @@ def get_qwen_tts():
     return _qwen_tts_model
 
 
-def get_custom_voice_model():
-    """Lazy-load the 0.6B CustomVoice model for stock voice. Thread-safe."""
-    global _custom_voice_model
-    if _custom_voice_model is None:
-        with _custom_model_lock:
-            if _custom_voice_model is None:
-                from qwen_tts import Qwen3TTSModel
-
-                device = "cuda" if torch.cuda.is_available() else "cpu"
-
-                logger.info("Loading CustomVoice 0.6B on %s...", device)
-                attn_impl = _select_attn_impl()
-                _custom_voice_model = Qwen3TTSModel.from_pretrained(
-                    CUSTOM_VOICE_MODEL_ID,
-                    device_map=device,
-                    attn_implementation=attn_impl,
-                )
-                logger.info("CustomVoice 0.6B loaded (attn=%s).", attn_impl)
-    return _custom_voice_model
-
-
 def _try_torch_compile(wrapper):
     """Best-effort torch.compile on model submodules. Disabled for stability."""
     # torch.compile can cause CUDA asserts on some GPU architectures (T4/Turing)
@@ -328,21 +306,13 @@ def load_default_profile() -> str | None:
 # Core API
 # ---------------------------------------------------------------------------
 
-def create_voice_profile(ref_audio_path: str, voice_name: str = "Cloned Voice") -> str:
+def create_voice_profile(ref_audio_path: str, voice_name: str = "Cloned Voice", profile_id_override: str | None = None) -> str:
     """
     Extract speaker embedding from reference audio, cache it, and save to disk.
     Returns a profile_id string for later synthesis.
 
     Reference audio is trimmed to 3-10s for optimal latency.
-    Only supported with the Base model (1.7B).
     """
-    if _MODEL_MODE == "custom_voice":
-        raise ValueError(
-            "Voice cloning requires the Base model (1.7B). "
-            "The 0.6B CustomVoice model only supports predefined speakers. "
-            "Set QWEN_TTS_MODE=base or remove the env var."
-        )
-
     trimmed_path = _trim_reference_audio(ref_audio_path)
     model = get_qwen_tts()
 
@@ -353,7 +323,7 @@ def create_voice_profile(ref_audio_path: str, voice_name: str = "Cloned Voice")
             x_vector_only_mode=True,
         )
 
-    profile_id = uuid.uuid4().hex[:12]
+    profile_id = profile_id_override or uuid.uuid4().hex[:12]
     with _cache_lock:
         _PROFILE_CACHE[profile_id] = prompt_items
 
@@ -397,26 +367,44 @@ def synthesize_cloned(text: str, profile_id: str) -> tuple[np.ndarray, int]:
     return wav, sample_rate
 
 
+def _ensure_stock_voice_profile():
+    """Ensure the stock vivian voice profile is loaded (uses Base 1.7B with reference audio)."""
+    with _cache_lock:
+        if STOCK_VOICE_PROFILE_ID in _PROFILE_CACHE:
+            return
+    # Create profile from vivian reference audio
+    if VIVIAN_REF_PATH.exists():
+        try:
+            create_voice_profile(str(VIVIAN_REF_PATH), voice_name="Vivian (Stock)", profile_id_override=STOCK_VOICE_PROFILE_ID)
+            logger.info("Stock vivian voice profile created from reference audio.")
+        except Exception as e:
+            logger.warning("Failed to create stock voice profile: %s", e)
+
+
 def synthesize_custom_voice_streaming(
     text: str, speaker: str = "vivian", language: str = "english"
 ):
     """
-    Synthesize text with the 0.6B CustomVoice model, yielding partial audio segments
-    as they become available for lower latency.
+    Synthesize text with the Base 1.7B model using stock vivian reference.
     Yields (wav_segment, sample_rate) tuples.
     """
-    model = get_custom_voice_model()
+    _ensure_stock_voice_profile()
+    with _cache_lock:
+        prompt_items = _PROFILE_CACHE.get(STOCK_VOICE_PROFILE_ID)
+    if prompt_items is None:
+        logger.warning("Stock voice profile not available, synthesizing without clone.")
+        return
 
+    model = get_qwen_tts()
     with GPU_INFERENCE_LOCK:
-        audio_list, sample_rate = model.generate_custom_voice(
+        audio_list, sample_rate = model.generate_voice_clone(
             text=text,
-            speaker=speaker,
             language=language,
+            voice_clone_prompt=prompt_items,
             non_streaming_mode=False,
             **GENERATION_PARAMS,
         )
 
-    # Yield each segment as it was generated
     for segment in audio_list:
         if segment is not None and len(segment) > 0:
             yield segment, sample_rate
@@ -426,23 +414,16 @@ def synthesize_custom_voice(
     text: str, speaker: str = "vivian", language: str = "english"
 ) -> tuple[np.ndarray, int]:
     """
-    Synthesize text with the 0.6B CustomVoice model (predefined speakers).
-    Much faster than cloned synthesis but no voice cloning.
+    Synthesize text with the Base 1.7B model using stock vivian reference.
     Returns (wav_array, sample_rate).
     """
-    model = get_custom_voice_model()
-
-    with GPU_INFERENCE_LOCK:
-        audio_list, sample_rate = model.generate_custom_voice(
-            text=text,
-            speaker=speaker,
-            language=language,
-            non_streaming_mode=False,
-            **GENERATION_PARAMS,
-        )
-
-    wav = np.concatenate(audio_list) if audio_list else np.zeros(0, dtype=np.float32)
-    return wav, sample_rate
+    segments = list(synthesize_custom_voice_streaming(text, speaker, language))
+    if segments:
+        wav = np.concatenate([s for s, _ in segments])
+        sr = segments[0][1]
+        return wav, sr
+    # Fallback: empty audio
+    return np.zeros(0, dtype=np.float32), 24000
 
 
 def synthesize_cloned_preview(profile_id: str) -> tuple[np.ndarray, int]:
@@ -465,6 +446,3 @@ def has_profile(profile_id: str | None) -> bool:
     return pt_path.exists()
 
 
-def get_model_mode() -> str:
-    """Return the current model mode ('base' or 'custom_voice')."""
-    return _MODEL_MODE