fix_denose_diar

main_pipeline.py

@@ -1,4 +1,5 @@
 import librosa
+import torch
 
 from utils.denoise_pipeline import denoise
 from utils.diarization_pipeline import diarization
@@ -32,11 +33,11 @@ def save_speaker_audios(segments, denoised_audio_path, out_folder='out', out_f=4
 
 
 def main_pipeline(audio_path):
-    denoised_audio_path = denoise(audio_path)
-    # denoised_audio_path = audio_path
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+    denoised_audio_path = denoise(audio_path, device)
     segments = diarization(denoised_audio_path)
     result_diarization = save_speaker_audios(segments, denoised_audio_path)
-    # return [denoised_audio_path] + result
     return denoised_audio_path, result_diarization
 
 

requirements.txt

@@ -19,4 +19,5 @@ lightning_fabric
 modelscope
 rotary_embedding_torch
 simple-diarizer
-soundfile
+soundfile
+librosa

utils/denoise_pipeline.py

@@ -23,8 +23,8 @@ def denoise(filename, device):
     vocal_wav = vocal_wav.numpy()
     vocal_wav = librosa.to_mono(vocal_wav)
     vocal_wav = vocal_wav.T
-    # vocal_wav = librosa.resample(vocal_wav, orig_sr=sr, target_sr=out_f)
-    write('denoise.wav', 44100, vocal_wav)
+    vocal_wav = librosa.resample(vocal_wav, orig_sr=44100, target_sr=48000)
+    write('denoise.wav', 48000, vocal_wav)
 
     return 'denoise.wav'
 

utils/diarization_pipeline.py

@@ -6,7 +6,7 @@ import soundfile as sf
 
 
 class DiarizationPipeline:
-    def __init__(self, mode='torch'):
+    def __init__(self, ):
         super(DiarizationPipeline, self).__init__()
         self.diar = Diarizer(
             embed_model='ecapa',  # supported types: ['xvec', 'ecapa']
@@ -26,7 +26,7 @@ class DiarizationPipeline:
         return segments
 
 
-diarization = DiarizationPipeline('torch')
+diarization = DiarizationPipeline()
 
 if __name__ == '__main__':
     diarization('../out.wav')