fma-small-encodec / README.md
hugfaceguy0001's picture
Update README.md
8e98887 verified
metadata
license: cc
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: prompt
      dtype: string
    - name: audio_code
      dtype:
        array2_d:
          shape:
            - null
            - 8
          dtype: int16
  splits:
    - name: train
      num_bytes: 357436897
      num_examples: 7916
  download_size: 180406553
  dataset_size: 357436897
task_categories:
  - text-to-audio
  - audio-classification
language:
  - en
tags:
  - music
  - art
pretty_name: fmaEncodec
size_categories:
  - 1K<n<10K

Free Music Archive (FMA) small encodec version

This dataset is built from Free Music Archive (FMA) small version ( https://huggingface.co/datasets/benjamin-paine/free-music-archive-small ). We combine the metadata in the original dataset into prompt, and use Encodec-24kHZ (https://huggingface.co/facebook/encodec_24khz) to encode the 30s audio into 2d array of shape (8,2249) or (8,2251) or the like.

  • "prompt": a description of the music, including title, author, genres, etc.
  • "audio_code": 2d array of shape [T, 8], T≈2249, indices are integers from 0 to 1023, encoded by encodec_24khz