README.md

metadata

dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          decode: false
    - name: duration
      dtype: float64
    - name: speakers
      dtype: string
    - name: source
      dtype: string
    - name: gender
      dtype: string
    - name: keep
      dtype: bool
    - name: duplicate
      dtype: bool
    - name: reasons
      list: string
    - name: clipped
      dtype: bool
    - name: silence_ratio
      dtype: float64
    - name: cps
      dtype: float64
    - name: word_count
      dtype: int64
    - name: detected_lang
      dtype: string
    - name: id
      dtype: string
    - name: text
      dtype: string
    - name: source_shard
      dtype: string
    - name: source_key
      dtype: string
  splits:
    - name: train
      num_bytes: 8643816083
      num_examples: 89055
  download_size: 8350172214
  dataset_size: 8643816083
language:
  - mos
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - automatic-speech-recognition
  - text-to-speech
size_categories:
  - 10K<n<100K

Clean Speech Dataset

As the name indicates, this dataset contains the rows retained by our curation pipeline.

The validation criteria are:

• Audio duration is within the accepted range
• Audio quality is sufficiently clear
• Noise levels are within the acceptable threshold

This clean dataset is ready for the model training.