metadata

license: cc0-1.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: esd.list
    column_names:
      - file_name
      - speaker
      - language
      - transcription
language:
  - yue
tags:
  - cantonese
  - audio
dataset_info:
  features:
    - name: file_name
      dtype: audio
    - name: speaker
      dtype: string
    - name: language
      dtype: string
    - name: transcription
      dtype: string

張悦楷三國演義

Fork from laubonghaudoi/zoengjyutgaai_saamgwokjinji

We found the original wav files are not splitted correctly, so we asked the author to provide the srt file and un-splitted wav files. We then re-split the wav files and align the srt file to the wav files. We also filtered some samples that are too short.

subtitles = []
splits = librosa.effects.split(audio) # shape: (682, 2)

!mkdir -p dataset/zoengjyutgaai_saamgwokjinji/wavs

# split audio by srt time
for i, sub in enumerate(subs):
    chunk_start = sub.start.to_time()
    chunk_end = sub.end.to_time()
    chunk_start = ((chunk_start.minute * 60) + chunk_start.second) * sr
    chunk_end = ((chunk_end.minute * 60) + chunk_end.second) * sr

    # Find the closest split
    chunk_start = min(splits, key=lambda x: abs(x[0] - chunk_start))[0]
    chunk_end = min(splits, key=lambda x: abs(x[1] - chunk_end))[1]

    chunk = audio[chunk_start:chunk_end]
    wav_file = f"001_{i:03}.wav"

    # resample, since bert-vits2 training only support 44.1k 
    try:
        chunk = librosa.resample(chunk, sr, 44100)
    except:
        print(f"Error resampling {wav_file}")
        continue

    subtitles.append({ 'path': wav_file, 'speaker': 'zoengjyutgaai', 'language': 'YUE', 'text': sub.text })

    # export audio
    sf.write(f"dataset/zoengjyutgaai_saamgwokjinji/wavs/{wav_file}", chunk, 44100, subtype='PCM_16')

df = pd.DataFrame(subtitles)
df.to_csv("dataset/zoengjyutgaai_saamgwokjinji/001.csv", index=False, sep='|', header=False)