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