I need to upload an audio file where two or more speakers are having a conversation, and at times their speech overlaps. The requirement is to segment the audio into distinct chunks, each corresponding clearly to a different speaker(Which speaker say what). I have used speaker diarization for this task, which includes pitch as a parameter for distinguishing speakers. However, when speakers have a similar pitch, the model fails to separate them correctly and treats them as the same person. What additional techniques or features can I incorporate to improve the diarization accuracy and handle overlapping speech more effectively? Here is my code: def run_speaker_diarization(audio_path): print("inside speaker diarization method") pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token="huggingface_token") diarization = pipeline(audio_path) audio = AudioSegment.from_wav(audio_path) # Output values wav_splits = [] labels = [] # Optional: save chunks base_dir = os.path.dirname(audio_path) base_name = os.path.splitext(os.path.basename(audio_path))[0] output_dir = os.path.join(base_dir, base_name + "_sliced_by_pyannote") os.makedirs(output_dir, exist_ok=True) for i, turn in enumerate(diarization.itertracks(yield_label=True)): segment, _, speaker = turn start = segment.start end = segment.end wav_splits.append((start, end)) # In seconds labels.append(speaker) # Optional: save chunk to file chunk = audio[int(start * 1000):int(end * 1000)] chunk_filename = f"{base_name}_Speaker{speaker}_chunk{i+1}.wav" chunk_path = os.path.join(output_dir, chunk_filename) chunk.export(chunk_path, format="wav") print(f" Saved: {chunk_filename}") print(f"\n All speaker segments saved to: {output_dir}") Continue reading...
