Load audio as waveform dict for pyannote (torchcodec unavailable on ROCm)

backend/app/services/diarization.py

@@ -25,8 +25,19 @@ def get_diarization_pipeline() -> Pipeline:
 
 
 def diarize(audio_path: str) -> list[dict]:
+    import torch
+    import soundfile as sf
+
     pipeline = get_diarization_pipeline()
-    diarization = pipeline(audio_path)
+
+    waveform, sample_rate = sf.read(audio_path, dtype="float32")
+    if waveform.ndim > 1:
+        waveform = waveform.mean(axis=1)
+    audio_input = {
+        "waveform": torch.from_numpy(waveform).unsqueeze(0),
+        "sample_rate": sample_rate,
+    }
+    diarization = pipeline(audio_input)
 
     segments = []
     for turn, _, speaker in diarization.itertracks(yield_label=True):