replace torchaudio.load() with sf.read()

unison/pipelines/infer.py

@@ -21,7 +21,6 @@ from typing import List, Optional
 import torch
 import torch.nn.functional as F
 import soundfile as sf
-import torchaudio
 import torchaudio.functional as AF
 from diffusers import FlowMatchEulerDiscreteScheduler
 from diffusers.training_utils import EMAModel
@@ -346,10 +345,19 @@ def _rms_normalize(wav: torch.Tensor, target_rms: float = TRAINING_TARGET_RMS) -
     return wav
 
 
+def _load_audio(audio_path):
+    """
+    Read an audio file into a [channels, samples] float32 tensor.
+    """
+    data, sr = sf.read(audio_path, dtype="float32", always_2d=True)  # [samples, channels]
+    wav = torch.from_numpy(data.T)  # [channels, samples]
+    return wav, sr
+
+
 def load_source_audio(audio_path, target_sr=DEFAULT_TARGET_SAMPLE_RATE,
                       target_length=None, max_duration=MAX_AUDIO_DURATION, device=None):
     device = device or torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    wav, sr = torchaudio.load(audio_path)
+    wav, sr = _load_audio(audio_path)
     if wav.shape[0] > 1:
         wav = wav.mean(0, keepdim=True)
     if sr != target_sr:
@@ -519,7 +527,7 @@ def load_ref_audio(
     the ref's last formant.
     """
     device = device or torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    wav, sr = torchaudio.load(audio_path)
+    wav, sr = _load_audio(audio_path)
     if wav.shape[0] > 1:
         wav = wav.mean(0, keepdim=True)
     if sr != target_sr: