fix: bypass torchaudio backend dispatch for torchaudio 2.9+

torchaudio.load now routes through load_with_torchcodec and ignores the
backend= kwarg, so the previous soundfile-backend patch failed with
'TorchCodec is required'. Replace torchaudio.load / torchaudio.info with
direct soundfile-based implementations to bypass backend dispatch
entirely. Falls back to librosa for formats soundfile cannot decode.

app.py

@@ -3,29 +3,66 @@ import gradio as gr
 import torch
 import librosa
 import numpy as np
-
-# Patch torchaudio to use soundfile backend globally.
-# PyTorch 2.9+ defaults to torchcodec which requires FFmpeg shared libs
-# that are incompatible with the HF Space environment.
+import soundfile as sf
+
+# Replace torchaudio.load / torchaudio.info with soundfile-backed versions.
+# Why: torchaudio 2.9+ routes torchaudio.load through load_with_torchcodec and
+# ignores the legacy backend= kwarg. torchcodec is unavailable in this Space,
+# so any indirect torchaudio.load call (inference.py, dataset_f.py, preprocess.py)
+# raises "TorchCodec is required for load_with_torchcodec". Bypassing the
+# backend dispatch entirely is the only stable fix.
 import torchaudio
 
-_orig_torchaudio_load = torchaudio.load
 
+def _patched_load(filepath, *args, **kwargs):
+    frame_offset = kwargs.pop("frame_offset", 0)
+    num_frames = kwargs.pop("num_frames", -1)
+    if len(args) >= 1:
+        frame_offset = args[0]
+    if len(args) >= 2:
+        num_frames = args[1]
 
-def _patched_load(*args, **kwargs):
-    kwargs.setdefault("backend", "soundfile")
-    return _orig_torchaudio_load(*args, **kwargs)
+    try:
+        data, sample_rate = sf.read(
+            str(filepath),
+            start=int(frame_offset) if frame_offset else 0,
+            frames=int(num_frames) if num_frames and num_frames > 0 else -1,
+            dtype="float32",
+            always_2d=True,
+        )
+        waveform = torch.from_numpy(data.T).contiguous()
+        return waveform, sample_rate
+    except Exception:
+        data, sample_rate = librosa.load(str(filepath), sr=None, mono=False)
+        if data.ndim == 1:
+            data = data[np.newaxis, :]
+        if frame_offset:
+            data = data[:, int(frame_offset):]
+        if num_frames and num_frames > 0:
+            data = data[:, : int(num_frames)]
+        waveform = torch.from_numpy(np.ascontiguousarray(data)).float()
+        return waveform, sample_rate
 
 
 torchaudio.load = _patched_load
-if hasattr(torchaudio, "info"):
-    _orig_torchaudio_info = torchaudio.info
 
-    def _patched_info(*args, **kwargs):
-        kwargs.setdefault("backend", "soundfile")
-        return _orig_torchaudio_info(*args, **kwargs)
 
-    torchaudio.info = _patched_info
+class _AudioInfo:
+    __slots__ = ("sample_rate", "num_frames", "num_channels", "bits_per_sample", "encoding")
+
+
+def _patched_info(filepath, *args, **kwargs):
+    info = sf.info(str(filepath))
+    out = _AudioInfo()
+    out.sample_rate = info.samplerate
+    out.num_frames = info.frames
+    out.num_channels = info.channels
+    out.bits_per_sample = 0
+    out.encoding = info.format
+    return out
+
+
+torchaudio.info = _patched_info
 
 from inference import inference
 from huggingface_hub import hf_hub_download