intronhealth/afrispeech-200

Dataset Card for AfriSpeech-200

Table of Contents

• —Dataset Card for AfriSpeech-200
• —Table of Contents
• —Dataset Description
• —Dataset Summary
• —How to use
• —Supported Tasks and Leaderboards
• —Languages
• —Dataset Structure
• —Data Instances
• —Data Fields
• —Data Splits
• —Dataset Creation
• —Curation Rationale
• —Source Data
• —Initial Data Collection and Normalization
• —Who are the source language producers?
• —Annotations
• —Annotation process
• —Who are the annotators?
• —Personal and Sensitive Information
• —Considerations for Using the Data
• —Social Impact of Dataset
• —Discussion of Biases
• —Other Known Limitations
• —Additional Information
• —Dataset Curators
• —Licensing Information
• —Citation Information
• —Contributions

Dataset Description

• —Homepage: https://github.com/intron-innovation/AfriSpeech-Dataset-Paper
• —Repository: https://github.com/intron-innovation/AfriSpeech-Dataset-Paper
• —Paper: AfriSpeech-200: Pan-African accented speech dataset for clinical and general domain ASR
• —Leaderboard: [Needs More Information]
• —Point of Contact: Intron Innovation

Dataset Summary

AFRISPEECH-200 is a 200hr Pan-African speech corpus for clinical and general domain English accented ASR; a dataset with 120 African accents from 13 countries and 2,463 unique African speakers. Our goal is to raise awareness for and advance Pan-African English ASR research, especially for the clinical domain.

How to use

The datasets library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the load_dataset function.

from datasets import load_dataset

afrispeech = load_dataset("tobiolatunji/afrispeech-200", "all")

The entire dataset is ~120GB and may take about 2hrs to download depending on internet speed/bandwidth. If you have disk space or bandwidth limitations, you can use streaming mode described below to work with smaller subsets of the data.

Alterntively you are able to pass a config to the load_dataset function and download only a subset of the data corresponding to a specific accent of interest. The example provided below is isizulu.

For example, to download the isizulu config, simply specify the corresponding accent config name. The list of supported accents is provided in the accent list section below:

from datasets import load_dataset

afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train")

Using the datasets library, you can also stream the dataset on-the-fly by adding a streaming=True argument to the load_dataset function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.

from datasets import load_dataset

afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train", streaming=True)

print(next(iter(afrispeech)))
print(list(afrispeech.take(5)))

Local

from datasets import load_dataset
from torch.utils.data.sampler import BatchSampler, RandomSampler

afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train")
batch_sampler = BatchSampler(RandomSampler(afrispeech), batch_size=32, drop_last=False)
dataloader = DataLoader(afrispeech, batch_sampler=batch_sampler)

Streaming

from datasets import load_dataset
from torch.utils.data import DataLoader

afrispeech = load_dataset("tobiolatunji/afrispeech-200", "isizulu", split="train", streaming=True)
dataloader = DataLoader(afrispeech, batch_size=32)

Caveats

Note that till the end of the ongoing AfriSpeech ASR Challenge event (Feb - May 2023), the transcripts in the validation set are hidden and the test set will be unreleased till May 19, 2023.

Fine-tuning Colab tutorial

To walk through a complete colab tutorial that finetunes a wav2vec2 model on the afrispeech-200 dataset with transformers, take a look at this colab notebook afrispeech/wav2vec2-colab-tutorial.

Supported Tasks and Leaderboards

• —Automatic Speech Recognition
• —Speech Synthesis (Text-to-Speech)

Languages

English (Accented)

Dataset Structure

Data Instances

A typical data point comprises the path to the audio file, called path and its transcription, called transcript. Some additional information about the speaker is provided.

{
    'speaker_id': 'b545a4ca235a7b72688a1c0b3eb6bde6', 
    'path': 'aad9bd69-7ca0-4db1-b650-1eeea17a0153/5dcb6ee086e392376cd3b7131a250397.wav', 
    'audio_id': 'aad9bd69-7ca0-4db1-b650-1eeea17a0153/5dcb6ee086e392376cd3b7131a250397',
    'audio': {
        'path': 'aad9bd69-7ca0-4db1-b650-1eeea17a0153/5dcb6ee086e392376cd3b7131a250397.wav', 
        'array': array([0.00018311, 0.00061035, 0.00012207, ..., 0.00192261, 0.00195312, 0.00216675]), 
        'sampling_rate': 44100}, 
    'transcript': 'His mother is in her 50 s and has hypertension .', 
    'age_group': '26-40', 
    'gender': 'Male', 
    'accent': 'yoruba', 
    'domain': 'clinical', 
    'country': 'US', 
    'duration': 3.241995464852608
}

Data Fields

• —speaker_id: An id for which speaker (voice) made the recording

• —path: The path to the audio file

• —audio: A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].

• —transcript: The sentence the user was prompted to speak

Data Splits

The speech material has been subdivided into portions for train, dev, and test.

Speech was recorded in a quiet environment with high quality microphone, speakers were asked to read one sentence at a time.

• —Total Number of Unique Speakers: 2,463
• —Female/Male/Other Ratio: 57.11/42.41/0.48
• —Data was first split on speakers. Speakers in Train/Dev/Test do not cross partitions

Dataset Creation

Curation Rationale

Africa has a very low doctor-to-patient ratio. At very busy clinics, doctors could see 30+ patients per day-- a heavy patient burden compared with developed countries-- but productivity tools such as clinical automatic speech recognition (ASR) are lacking for these overworked clinicians. However, clinical ASR is mature, even ubiquitous, in developed nations, and clinician-reported performance of commercial clinical ASR systems is generally satisfactory. Furthermore, the recent performance of general domain ASR is approaching human accuracy. However, several gaps exist. Several publications have highlighted racial bias with speech-to-text algorithms and performance on minority accents lags significantly. To our knowledge, there is no publicly available research or benchmark on accented African clinical ASR, and speech data is non-existent for the majority of African accents. We release AfriSpeech, 200hrs of Pan-African speech, 67,577 clips from 2,463 unique speakers, across 120 indigenous accents from 13 countries for clinical and general domain ASR, a benchmark test set, with publicly available pre-trained models with SOTA performance on the AfriSpeech benchmark.

Source Data

Country Stats

Accent Stats

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

[Needs More Information]

Who are the annotators?

[Needs More Information]

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

Dataset provided for research purposes only. Please check dataset license for additional information.

Additional Information

Dataset Curators

The dataset was initially prepared by Intron and refined for public release by CLAIR Lab.

Licensing Information

Public Domain, Creative Commons Attribution NonCommercial ShareAlike v4.0 (CC BY-NC-SA 4.0)

Citation Information

@article{olatunji2023afrispeech, title={AfriSpeech-200: Pan-African Accented Speech Dataset for Clinical and General Domain ASR}, author={Olatunji, Tobi and Afonja, Tejumade and Yadavalli, Aditya and Emezue, Chris Chinenye and Singh, Sahib and Dossou, Bonaventure FP and Osuchukwu, Joanne and Osei, Salomey and Tonja, Atnafu Lambebo and Etori, Naome and others}, journal={arXiv preprint arXiv:2310.00274}, year={2023} }

Contributions

Thanks to @tobiolatunji for adding this dataset.