main_pipeline.py

import librosa
import torch
import os
from utils.denoise_pipeline import denoise
from utils.diarization_pipeline import diarization
import numpy as np

import pandas as pd
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 = []
        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'{out_folder}/{label}.wav'
        sf.write(out_wav_path, np.concatenate(output_signal), out_f, 'PCM_24')

        out_wav_paths.append(out_wav_path)

    return out_wav_paths


def main_pipeline(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_diarization


if __name__ == '__main__':
    main_pipeline('dialog.mp3')