README.md

metadata

dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: sentence
      dtype: string
  splits:
    - name: train
      num_bytes: 12290918193
      num_examples: 42566
  download_size: 10107309830
  dataset_size: 12290918193
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: apache-2.0
task_categories:
  - automatic-speech-recognition
language:
  - ja
tags:
  - audio
  - nlp
  - asr
size_categories:
  - 10K<n<100K

   
dataset = load_dataset("sin2piusc/CSS10_jsss_jvs_cv")
output_file = 'metadata.csv'

special_characters = '[,♬「」 ？！;:“%‘” ~♪…～ ！?!？#$%&゛＃＄％＆（）＊＋：；〈＝〉？＠＾＿‘｛｜｝～".,]'

for i, sample in tqdm(enumerate(dataset)): # Process each sample in the filtered dataset
    audio_sample = name + f'_{i}.mp3'
    audio_path = os.path.join(folder_path, audio_sample)
    transcription_path = os.path.join(folder_path, output_file)  # Path to save transcription file 
    sf.write(audio_path, sample['audio']['array'], sample['audio']['sampling_rate']) # Save mp3/wav files
    sample["audio_length"] = len(sample["audio"]["array"]) / sample["audio"]["sampling_rate"]  # Get audio length
    with open(transcription_path, 'a', encoding='utf-8') as transcription_file: # Save transcription file    
        sample["sentence"] = re.sub(special_characters,'', sample["sentence"]) # Remove special characters 
        transcription_file.write(audio_sample + ",") # Save transcription file name
        transcription_file.write(sample['sentence']) # Save transcription 
        transcription_file.write(str(","+str(sample['audio_length']))) # Save audio length
        transcription_file.write('\n')