Updated by Bas Curtiz. Seems to work better this way!

demucs_stem_organizer_ui.py

@@ -106,6 +106,31 @@ def next_sequence_number(out_dir: Path, manifest: dict) -> int:
 
 
 def load_audio(path: str, sr: int, ch: int = 2) -> np.ndarray:
+    p = Path(path)
+    sf_exts = {'.wav', '.flac', '.aif', '.aiff', '.ogg'}
+    if p.suffix.lower() in sf_exts:
+        try:
+            data, file_sr = sf.read(str(p), dtype='float32', always_2d=True)
+            # sf.read returns (samples, channels); convert to (channels, samples)
+            audio = data.T
+            # Normalise channel count
+            if audio.shape[0] == 1:
+                audio = np.repeat(audio, ch, axis=0)
+            elif audio.shape[0] > ch:
+                audio = audio[:ch]
+            # Resample if needed
+            if file_sr != sr:
+                try:
+                    import resampy
+                    audio = resampy.resample(audio, file_sr, sr, axis=1)
+                except ImportError:
+                    raise RuntimeError(
+                        f"Sample rate mismatch ({file_sr} Hz vs expected {sr} Hz) "
+                        "and resampy is not installed. Install it with: pip install resampy"
+                    )
+            return audio.astype(np.float32)
+        except Exception:
+            pass  # fall through to AudioFile / ffmpeg
     return AudioFile(path).read(streams=0, samplerate=sr, channels=ch).numpy().astype(np.float32)