juliusのセグメンテーション結果に基づいて音声をトリミングする

trim.py

import numpy as np
import os
from nnmnkwii.datasets import jsut

import librosa
import librosa.display
from matplotlib import pyplot as plt
from nnmnkwii.io import hts
from os.path import exists

if __name__ == "__main__":
    in_dir = "/home/ryuichi/data/jsut_ver1"
    transcriptions = jsut.TranscriptionDataSource(
        in_dir, subsets=jsut.available_subsets).collect_files()
    wav_paths = jsut.WavFileDataSource(
        in_dir, subsets=jsut.available_subsets).collect_files()

    for subset in jsut.available_subsets:
        save_dir = join(in_dir, subset, "wav_trim")
        os.makedirs(save_dir, exist_ok=True)
 
    for idx, (text, wav_path) in enumerate(zip(transcriptions, wav_paths)):
        print(idx, wav_path)
        lab_path = wav_path.replace("wav/", "lab/").replace(".wav", ".lab")

        x, sr = librosa.load(wav_path, sr=20000)
        y, _ = librosa.effects.trim(x, top_db=30)

        if exists(lab_path):
            labels = hts.load(lab_path)
            assert labels[0][-1] == "silB"
            assert labels[-1][-1] == "silE"
            b = int(labels[0][1] * 1e-7 * sr)
            e = int(labels[-1][0] * 1e-7 * sr)
            yy = x[b:e]
        else:
            yy = x
            
        out_wav_path = wav_path.replace("wav/", "wav_trim/")
        librosa.output.write_wav(out_wav_path, yy, sr)