Implementation:Facebookresearch Audiocraft InfoAudioDataset

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Audio_Data, Dataset
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for loading audio datasets with metadata annotations returned alongside waveforms as SegmentWithAttributes provided by the AudioCraft library.

Description

InfoAudioDataset extends AudioDataset to ensure that every sample returned includes metadata (text descriptions, artist info, etc.) as a SegmentWithAttributes alongside the audio tensor. It applies cluster-specific path remapping on initialization and defines a base AudioInfo dataclass for simple metadata containers.

Usage

Import this class as the base for metadata-aware audio datasets such as MusicDataset, SoundDataset, and JascoDataset.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/data/info_audio_dataset.py
• Lines: 1-110

Signature

class InfoAudioDataset(AudioDataset):
    def __init__(self, meta: tp.List[AudioMeta], **kwargs):
        """Applies clusterify_all_meta and forwards to AudioDataset."""

    def __getitem__(self, index: int):
        """Returns (wav, AudioInfo) tuple."""

Import

from audiocraft.data.info_audio_dataset import InfoAudioDataset, AudioInfo

I/O Contract

Inputs

Name	Type	Required	Description
meta	List[AudioMeta]	Yes	List of audio file metadata entries
index	int	Yes	Sample index

Outputs

Name	Type	Description
wav	torch.Tensor	Audio waveform [C, T]
info	AudioInfo	Metadata segment with optional audio_tokens

Usage Examples

from audiocraft.data.info_audio_dataset import InfoAudioDataset

# Typically used as a base class:
# class SoundDataset(InfoAudioDataset): ...
# class MusicDataset(InfoAudioDataset): ...

Related Pages

• Principle:Facebookresearch_Audiocraft_Audio_Dataset_Preparation