关于唱法模型数据集

[lhc991025]

要训练唱法模型，可以用公开的opencpop数据集吗？但是它里面的transcriptions.txt文件和你所描述的transcriptions.csv文件不太符合，所以是不能直接用opencpop数据集进行训练吗？

[yqzhishen]

opencpop的原始标注格式有问题，所以要用也得用转换过的。

但依然不建议用它训练，因为它的标注质量不咋地

[contiunity]

不建议用opencpop本身的标注训练，要用opencpop建议自己重新标注

[LingYi0110]

如果您坚持用opencpop原始标注的话，这个能帮助你转换一部分 : )

import os

def parse_line(line):
    parts = line.split('|')
    return {
        'name': parts[0],
        'ph_seq': parts[2].split(' '),
        'ph_dur': parts[-2].split(' '),
        'slur': parts[-1].split(' ')
    }

def handle_slur(wav):
    for i, slur_val in enumerate(wav['slur']):
        if slur_val == '1':
            wav['ph_seq'][i] = ''
            offset = 0
            while True:
                if wav['slur'][i - offset] == '0':
                    wav['ph_dur'][i - offset] = str(
                        float(wav['ph_dur'][i - offset]) + float(wav['ph_dur'][i])
                    )
                    wav['ph_dur'][i] = ''
                    break
                offset += 1

def cleanup(wav, key):
    wav[key] = [x for x in wav[key] if x]

def apply_map(ph_seq, mapping):
    i = 0
    while i < len(ph_seq) - 1:
        pair = (ph_seq[i], ph_seq[i+1])
        for old, new in mapping.items():
            old_pair = tuple(old.split())
            new_pair = new.split()
            if pair == old_pair:
                ph_seq[i], ph_seq[i+1] = new_pair
        i += 1

def main():
    transcription_in = os.path.join('path/to/your/opencpop/segments', 'transcriptions.txt')
    transcription_out = os.path.join('/path/to/your/save', 'transcriptions.csv')

    mapping = {
        'ch i': 'ch ir',
        'c i': 'c i0',
        'r i': 'r ir',
        'sh i': 'sh ir',
        's i': 's i0',
        'y an': 'y En',
        'y e': 'y E',
        'zh i': 'zh ir',
        'z i': 'z i0',
    }

    segments = []
    with open(transcription_in, 'r', encoding='utf8') as f:
        for line in f:
            wav = parse_line(line.strip())
            handle_slur(wav)
            cleanup(wav, 'ph_seq')
            cleanup(wav, 'ph_dur')
            apply_map(wav['ph_seq'], mapping)
            segments.append([
                wav['name'],
                ' '.join(wav['ph_seq']),
                ' '.join(wav['ph_dur'])
            ])

    with open(transcription_out, 'w', encoding='utf8') as f:
        f.write('name,ph_seq,ph_dur\n')
        for name, ph_seq, ph_dur in segments:
            f.write(f'{name},{ph_seq},{ph_dur}\n')

if __name__ == '__main__':
    main()