anforsm's picture
Update README.md
a64784e
metadata
license: cc0-1.0
language:
  - en
task_categories:
  - text-to-speech
  - text-generation
pretty_name: Common Voice 11 (en) Cleaned and Tokenized
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: input_ids
      sequence: int32
    - name: attention_mask
      sequence: int8
    - name: labels
      sequence: int64
  splits:
    - name: train
      num_bytes: 1109542776
      num_examples: 83274
    - name: validation
      num_bytes: 17374496
      num_examples: 1304
  download_size: 197852035
  dataset_size: 1126917272

A cleaned and tokenized version of the English data from Mozilla Common Voice 11 dataset.

Cleaning steps:

  • Filtered on samples with >2 upvotes and <1 downvotes]
  • Removed non voice audio at start and end through pytorch VAD

Tokenization:

  • Audio tokenized through EnCodec by Meta
    • Using 24khz pre-trained model, and target bandwidth of 1.5
    • Represented in text as audio_token_0 - audio_token_1023
  • Prompts constructed as "text: <common voice transcript>\naudio: <audio tokens>"
  • Prompts tokenized with GPT tokenizer with added vocab of audio tokens.
  • Tokenized prompts padded to size 1024 with eos_token.

Each sample has 3 properties: input_ids, attention_mask and labels. input_ids and labels are the tokenized prompts and attention_mask is the attention mask.