Laravel

Introduction

🤗 Datasets is an open-source library for downloading and preparing datasets across all domains. Its minimalistic API lets you load a dataset in one line of Python, with efficient pre-processing functions. It offers a vast collection of audio datasets, making it a go-to tool for researchers and practitioners. This guide covers audio-specific features that streamline working with audio data.

The Hub

The Hugging Face Hub hosts open-source models, datasets, and demos. It includes a growing collection of audio datasets for various domains, tasks, and languages. Thanks to tight integration with 🤗 Datasets, any dataset on the Hub can be downloaded with a single line of code.

Filter datasets by task:

Currently, there are 77 speech recognition and 28 audio classification datasets, with numbers growing. Clicking on a dataset like common_voice reveals a dataset card with details, models trained on it, and a preview of audio samples you can play in real time.

Load an Audio Dataset

The load_dataset function handles downloading, extracting, and preparing data in one step. For example, to load the GigaSpeech dataset:

from datasets import load_dataset

gigaspeech = load_dataset("speechcolab/gigaspeech", "xs")
print(gigaspeech)

Output:

DatasetDict({
    train: Dataset({...})
    validation: Dataset({...})
    test: Dataset({...})
})

The returned DatasetDict acts like a Python dictionary. Access the training split with gigaspeech["train"] and the first item with gigaspeech["train"][0].

Easy to Load, Easy to Process

Audio datasets include an audio feature that stores the audio array, sampling rate, and path. You can cast audio columns to a desired sampling rate and apply transformations efficiently using map. The library also supports streaming mode, which loads data on the fly without downloading the entire dataset.

Streaming Mode: The Silver Bullet

Streaming mode is ideal for huge datasets that don't fit in memory. Enable it by passing streaming=True to load_dataset. Data is fetched in chunks, allowing you to iterate over samples without full download.

A Tour of Audio Datasets on the Hub

Explore datasets by task or language. Use the Hub's filters to find datasets like common_voice (multilingual speech recognition) or esc50 (audio classification). Each dataset card includes a preview, making it easy to assess quality before use.

Closing Remarks

🤗 Datasets simplifies audio dataset handling from download to preprocessing. With one-line loading, built-in audio features, and streaming, it's an essential tool for audio machine learning projects.

Your Ultimate Guide to Audio Datasets on Hugging Face