fix: audio decode normalization + warmup first 5 paragraphs

- inference_lfm.py: normalize waveform to [-1,1], handle dtype/device issues,
detect garbage audio (too quiet), proper float32 conversion
- qa_flow.py: write WAV as FLOAT subtype for quality
- app.py: warmup expanded from 1 to 5 paragraphs per story

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

app.py

@@ -423,28 +423,33 @@ print("[MomsVoice] LFM Q&A model ready. All models preloaded.")
 import threading
 
 def _warmup_first_paragraphs():
-    """Background: pre-generate first paragraph of each story with stock voice to disk cache."""
+    """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
-            first_chunks = split_into_chunks(paras[0])
-            for chunk in first_chunks:
+            # 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)
-            _logger.info("Warmed: %s (%d chunks)", book["title"], len(first_chunks))
+                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 first paragraphs warmed.")
+    _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 paragraph for all stories.")
+print("[MomsVoice] Background warmup started: pre-generating first 5 paragraphs for all stories.")
 
 
 # Gradio Application Core setup

assets/vivian_reference.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddd916e813a290c83be9aad1839f58739be153749d47a3e01d926b1cde26bf4e
+size 368684

inference_lfm.py

@@ -152,10 +152,28 @@ def answer_question_audio(
     # Decode audio
     waveform = None
     if audio_out and len(audio_out) > 1:
-        decode_device = _module_device(processor, _module_device(model, _select_device()))
-        audio_codes = torch.stack(audio_out[:-1], 1).unsqueeze(0).to(decode_device)
-        with GPU_INFERENCE_LOCK, torch.inference_mode():
-            waveform_tensor = processor.decode(audio_codes)
-        waveform = waveform_tensor.cpu().numpy().squeeze()
+        try:
+            decode_device = _module_device(processor, _module_device(model, _select_device()))
+            # Stack all audio codes except the final EOS marker
+            audio_codes = torch.stack(audio_out[:-1], 1).unsqueeze(0)
+            # Ensure codes are on the right device and in int/long format
+            if audio_codes.is_floating_point():
+                audio_codes = audio_codes.long()
+            audio_codes = audio_codes.to(decode_device)
+            with GPU_INFERENCE_LOCK, torch.inference_mode():
+                waveform_tensor = processor.decode(audio_codes)
+            waveform = waveform_tensor.detach().cpu().float().numpy().squeeze()
+            # Normalize to [-1, 1] range to prevent clipping/distortion
+            if waveform.ndim == 0:
+                waveform = None
+            elif len(waveform) > 0:
+                peak = np.abs(waveform).max()
+                if peak > 1.0:
+                    waveform = waveform / peak * 0.95
+                elif peak < 0.01:
+                    waveform = None  # Too quiet, likely garbage
+        except Exception as e:
+            logger.warning("Audio decode failed: %s", e)
+            waveform = None
 
     return answer_text, waveform, SAMPLE_RATE

qa_flow.py

@@ -68,8 +68,11 @@ def release_audio_submission(audio_path: str | Path | None) -> None:
 
 def _default_audio_writer(path: Path, waveform, sample_rate: int) -> None:
     import soundfile as sf
+    import numpy as np
 
-    sf.write(str(path), waveform, sample_rate)
+    # Ensure float32 for proper WAV encoding
+    wav = np.asarray(waveform, dtype=np.float32)
+    sf.write(str(path), wav, sample_rate, subtype='FLOAT')
 
 
 def _prune_generated_answers(output_dir: Path) -> None: