README.md

metadata

language:
  - uz
license: apache-2.0
size_categories:
  - 100K<n<1M
task_categories:
  - automatic-speech-recognition
dataset_info:
  features:
    - name: path
      dtype:
        audio:
          sampling_rate: 16000
    - name: text
      dtype: string
    - name: previous_text
      dtype: string
    - name: id
      dtype: int64
    - name: client_id
      dtype: string
    - name: duration
      dtype: float64
    - name: sentence
      dtype: string
    - name: created_at
      dtype: string
    - name: original_sentence_id
      dtype: string
    - name: sentence_clips_count
      dtype: int64
    - name: upvotes_count
      dtype: int64
    - name: downvotes_count
      dtype: int64
    - name: reported_count
      dtype: int64
    - name: reported_reasons
      dtype: string
    - name: skipped_clips
      dtype: int64
    - name: gender
      dtype: string
    - name: accent_region
      dtype: string
    - name: native_language
      dtype: string
    - name: year_of_birth
      dtype: string
  splits:
    - name: train
      num_bytes: 13791343519.24
      num_examples: 501330
    - name: validate
      num_bytes: 57649995.584
      num_examples: 2048
  download_size: 13680801049
  dataset_size: 13848993514.824
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validate
        path: data/validate-*

This is heavy filtered version of the dataset with additional information. This dataset does not contain original Mozilla Common Voice audios or texts

We filtered the dataset using number approaches:

1. VAD + Noise detection. Audios which lacked voice activity and produced no sound after denoiser were removed
2. Reading Speed. Audios with outlier speeds (approximately 5-10%), as they didnt match natural speed or were too noisy
3. Automatic STT validation. We trained the model using subset of valid samples from different authors and used trained model to extend the number of samples given their transcription match our trained model output to some extend, then we repeated this step multiple times until we reached this dataset size