Fix torchcodec issue

src/karaoke_maker/audio_io.py

@@ -0,0 +1,14 @@
+from pathlib import Path
+
+import soundfile as sf
+import torch
+
+
+def load_wav(path: Path) -> tuple[torch.Tensor, int]:
+    """Load a WAV file as a (channels, samples) float32 tensor.
+
+    Reads via soundfile rather than torchaudio.load, whose default backend
+    now requires the separate torchcodec package.
+    """
+    data, sr = sf.read(str(path), dtype="float32", always_2d=True)
+    return torch.from_numpy(data.T).contiguous(), sr

src/karaoke_maker/stages/align.py

@@ -23,7 +23,9 @@ def force_align_section(vocals_path: Path, start: float, end: float,
     import torchaudio
     from torchaudio.pipelines import MMS_FA as bundle
 
-    waveform, sr = torchaudio.load(str(vocals_path))
+    from ..audio_io import load_wav
+
+    waveform, sr = load_wav(vocals_path)
     waveform = waveform.mean(0, keepdim=True)
     lo, hi = int(start * sr), min(int(end * sr), waveform.shape[1])
     waveform = torchaudio.functional.resample(

src/karaoke_maker/stages/pitch.py

@@ -9,7 +9,9 @@ def track_pitch(vocals_path: Path, model: str, device: str):
     import torchaudio
     import torchcrepe
 
-    waveform, sr = torchaudio.load(str(vocals_path))
+    from ..audio_io import load_wav
+
+    waveform, sr = load_wav(vocals_path)
     waveform = waveform.mean(0, keepdim=True)
     waveform = torchaudio.functional.resample(waveform, sr, 16000)
     f0, periodicity = torchcrepe.predict(

tests/test_audio_io.py

@@ -0,0 +1,26 @@
+import numpy as np
+import soundfile as sf
+
+from karaoke_maker.audio_io import load_wav
+
+
+def test_load_wav_mono(tmp_path):
+    path = tmp_path / "mono.wav"
+    data = np.linspace(-0.5, 0.5, 1600, dtype=np.float32)
+    sf.write(path, data, 16000)
+    waveform, sr = load_wav(path)
+    assert sr == 16000
+    assert waveform.shape == (1, 1600)
+    assert np.allclose(waveform[0].numpy(), data, atol=1e-4)
+
+
+def test_load_wav_stereo(tmp_path):
+    path = tmp_path / "stereo.wav"
+    left = np.linspace(-0.5, 0.5, 800, dtype=np.float32)
+    right = np.linspace(0.5, -0.5, 800, dtype=np.float32)
+    sf.write(path, np.stack([left, right], axis=1), 16000)
+    waveform, sr = load_wav(path)
+    assert sr == 16000
+    assert waveform.shape == (2, 800)
+    assert np.allclose(waveform[0].numpy(), left, atol=1e-4)
+    assert np.allclose(waveform[1].numpy(), right, atol=1e-4)