education_chunked / README.md
instinct1912's picture
Update README.md
c59f97f verified
metadata
dataset_info:
  features:
    - name: transcription
      dtype: string
    - name: start
      dtype: float64
    - name: end
      dtype: float64
    - name: duration
      dtype: float64
    - name: word_count
      dtype: int64
    - name: source_file
      dtype: string
    - name: audio_quality
      struct:
        - name: rms_energy
          dtype: float64
        - name: silence_ratio
          dtype: float64
    - name: vad_refinement
      struct:
        - name: ctc_end
          dtype: float64
        - name: ctc_start
          dtype: float64
        - name: enabled
          dtype: bool
        - name: end_adjustment
          dtype: float64
        - name: search_windows
          sequence: float64
        - name: start_adjustment
          dtype: float64
        - name: vad_end
          dtype: float64
        - name: vad_start
          dtype: float64
    - name: speaker_validation
      struct:
        - name: has_overlap
          dtype: bool
        - name: is_valid
          dtype: bool
        - name: overlap_duration
          dtype: float64
        - name: rejection_reason
          dtype: 'null'
        - name: segments
          list:
            - name: end
              dtype: float64
            - name: speaker
              dtype: string
            - name: start
              dtype: float64
        - name: speaker_count
          dtype: int64
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_bytes: 214388341262.102
      num_examples: 1292013
  download_size: 208980676541
  dataset_size: 214388341262.102
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

TTS Audio Dataset

Description

Uzbek audio dataset for STT and TTS tasks.
All audio is cleaned and split into short chunks.
Each chunk has a transcription with word-level timestamps.


Data Format

  • Audio: FLAC, 24 kHz, 16-bit PCM
  • Text: Transcriptions with timestamps (UTF-8)

Source

Collected from publicly available materials.


Processing Notes

Audio is denoised and chunked. Text is aligned at the word level for training.


Use Cases

Ready for training speech-to-text and text-to-speech models.