README.md

metadata

dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: transcription
      dtype: string
  splits:
    - name: train
      num_bytes: 3125353264.6964455
      num_examples: 5778
    - name: test
      num_bytes: 1004055850.0756147
      num_examples: 1683
  download_size: 3490774262
  dataset_size: 4129409114.7720604
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
task_categories:
  - automatic-speech-recognition
  - text-to-speech
language:
  - km
tags:
  - openslr42
  - fleurs
  - asr

NOTE: If your colab crashes, please use pip install --upgrade --quiet datasets[audio]==3.6.0 to install datasets[audio] version 3.6.0.

This dataset combined google/fleurs, openslr/openslr42, and cleaned seanghay/khmer_mpwt_speech. Severals processes are executed:

1. clean up seanghay/khmer_mpwt_speech: manually correct wrong transcriptions over 2058 rows
2. normalize transcription: remove invisible white space; process ៗ, numbers, currencies, date into khmer text; and separate each word by space
3. filter out texts whose number of token ids are more than 448: use tokenizer of Whisper-Small to encode text and filter out sequences longer than 448
4. filter out audio with length longer than 30 seconds
5. resample audio to 16000kHz

Disclaimer I do not own any of these datasets.