metadata

license: mit
task_categories:
  - automatic-speech-recognition
language:
  - wo
tags:
  - ASR
  - Wolof
  - Audio
size_categories:
  - 10K<n<100K

Wolof Speech Dataset

This repository hosts a collection of Wolof speech datasets. The datasets, sourced from various contributors, are each stored in separate pickle files. Each pickle file contains the following columns:

audio: The audio data in WAV or MP3 format.
transcription: The corresponding transcriptions of the audio data.
length(duration(s)): The duration of each audio recording.
filename: The name of the audio file.
path: The path to the audio file.

Dataset Overview

To ensure well-organized and readable code, we have divided the code into separate files, each dedicated to downloading and cleaning a specific dataset. Below are the datasets included in this collection:

1. ISMA/ALFFA_WOLOF

Source: https://huggingface.co/datasets/Isma/alffa_wolof
Processing: A function was created to extract the audio from the column and save it in a WAV file.

2. SERGE-WILSON/WOLOF_SPEECH_TRANSCRIPTION

Source: Mendy, n.d.
Description: This dataset contains 14,844 audio rows or speech recordings, totaling 17.75 hours of recordings. The audio column type is 'ByteWAV'.
Processing: We implemented a function to convert the audio into WAV format before saving the corresponding files.

3. PERRYNELSON/WAXAL-WOLOF

Source: https://huggingface.co/datasets/perrynelson/waxal-wolof

4. GOOGLE/FLEURS

Source: Conneau et al., 2022
Description: This dataset is a comprehensive collection containing speech-to-text data for hundreds of languages, including low-resource languages like Wolof. It includes 11.2 hours of recordings for the Wolof dataset.

Other Website Sources

The following datasets were acquired by downloading a zip file containing folders with audio files and a file (CSV or TXT) housing transcriptions along with the paths to the corresponding audio files:

1. ALFFA DATASET

Source: Gauthier, Besacier, Voisin, Melese, & Elingui, 2016
Description: This dataset comprises various African languages designed for AI projects, aimed at leveraging low-resource African languages for AI applications. It contains 18.97 hours of Wolof recordings.

2. WAXAL DATASET

Source: Waxal-Multilingual, 2022
Description: This dataset provides a CSV file containing transcriptions and the paths to corresponding audio files, along with a link for downloading the audio files. It includes 7.12 hours of recordings.

3. WOLOF TTS DATASET

Source: Diop, Aw, Jaane, & Badiane, 2021
Description: The WOLOF TTS dataset presents a well-organized dataset categorized by gender voice. It is the most extensive dataset identified so far, comprising 45.3 hours of recordings.