fix: correct LFM audio sample rate to 48kHz and normalize volume

- SAMPLE_RATE 24000 -> 48000 (codec outputs 1920 samples/frame × 25 fps)
- Always normalize to 0.9 peak (was only scaling down, causing quiet output)
- Save answer WAV as PCM_16 for universal compatibility

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

inference_lfm.py

@@ -18,7 +18,7 @@ _model = None
 _model_lock = threading.Lock()
 
 HF_REPO = "LiquidAI/LFM2.5-Audio-1.5B"
-SAMPLE_RATE = 24000
+SAMPLE_RATE = 48000  # Codec outputs 1920 samples/frame at 25 frames/sec = 48kHz
 
 
 def _select_device() -> torch.device:
@@ -163,15 +163,15 @@ def answer_question_audio(
             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
+                if peak < 0.001:
+                    waveform = None  # Silent, likely garbage
+                else:
+                    # Always normalize to target peak (0.9) for consistent volume
+                    waveform = waveform * (0.9 / peak)
         except Exception as e:
             logger.warning("Audio decode failed: %s", e)
             waveform = None

qa_flow.py

@@ -70,9 +70,10 @@ def _default_audio_writer(path: Path, waveform, sample_rate: int) -> None:
     import soundfile as sf
     import numpy as np
 
-    # Ensure float32 for proper WAV encoding
     wav = np.asarray(waveform, dtype=np.float32)
-    sf.write(str(path), wav, sample_rate, subtype='FLOAT')
+    # Clip to [-1, 1] before writing as PCM_16 for universal compatibility
+    wav = np.clip(wav, -1.0, 1.0)
+    sf.write(str(path), wav, sample_rate, subtype='PCM_16')
 
 
 def _prune_generated_answers(output_dir: Path) -> None:

test_lfm_48k.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7af76a2ec139419c4f5ffdd4d0eb41b40ed0709faa1c3405541aa60969755935
+size 503084

test_lfm_local.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8237b2b464343d4b4f00a6e8e573f17cdf7589f1671903e0927bfd1429855670
+size 503084

test_lfm_normalized.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdbdc9d071ff8976805a1808df663754579807f6663762165d12cfdc47f0b4b9
+size 503084