denoise_diar_app

main_pipeline.py

@@ -1,6 +1,6 @@
 import librosa
 import torch
-
+import os
 from utils.denoise_pipeline import denoise
 from utils.diarization_pipeline import diarization
 import numpy as np
@@ -11,20 +11,20 @@ import soundfile as sf
 
 def save_speaker_audios(segments, denoised_audio_path, out_folder='out', out_f=48000):
     signal, sr = librosa.load(denoised_audio_path, sr=out_f, mono=True)
-
+    os.makedirs(out_folder, exist_ok=True)
     out_wav_paths = []
     segments = pd.DataFrame(segments)
 
     for label in set(segments.label):
         temp_df = segments[segments.label == label]
-        output_signal = np.zeros(int(len(signal)))
+        output_signal = []
         for _, r in temp_df.iterrows():
             start = int(r["start"] * out_f)
             end = int(r["end"] * out_f)
             output_signal[start:end] = signal[start:end]
 
-        out_wav_path = f'{label}.wav'
-        sf.write(out_wav_path, output_signal, out_f, 'PCM_24')
+        out_wav_path = f'{out_folder}/{label}.wav'
+        sf.write(out_wav_path, np.concatenate(output_signal), out_f, 'PCM_24')
 
         out_wav_paths.append(out_wav_path)
 

utils/diarization_pipeline.py

@@ -12,7 +12,7 @@ class DiarizationPipeline:
             embed_model='ecapa',  # supported types: ['xvec', 'ecapa']
             cluster_method='ahc',  # supported types: ['ahc', 'sc']
             window=1,  # size of window to extract embeddings (in seconds)
-            period=0.2  # hop of window (in seconds)
+            period=0.1  # hop of window (in seconds)
         )
 
     def __call__(self, wav_file):