AfriVox: An African benchmark dataset for Automatic Speech Translation and Speech Recognition

Project Overview

This project creates a benchmark dataset for evaluating Automatic Speech Translation and Speech recognition models on African languages. This benchmark dataset covers 18 African languages. See language details below.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Usage Instructions

Accessing the Dataset: The dataset can be accessed through Hugging Face:

from datasets import load_dataset
afrivox = load_dataset("intronhealth/afrivox")

Afrivox stats

• Dataset size = 18,881
• Total number of languages = 18
• Total number of hours = 64.26

Duration (hours) per domian:

domain	duration (hours)
medical	35.27
non-medical	28.99

Gender Distribution

gender	count
Male	8879
Female	10002

Dataset Summary by Language and Domain

language	medical (hrs)	non-medical (hrs)	total_hrs	num_speakers	num_samples
afrikaans	1.79	2.27	4.05	42	1406
akan	0.67	0.58	1.24	15	411
amharic	0.37	0.26	0.62	8	214
arabic	1.47	1.13	2.60	26	799
french	0.30	0.21	0.51	9	135
ga	0.00	0.01	0.01	1	5
hausa	3.81	1.71	5.53	125	1869
igbo	1.66	1.42	3.08	37	970
kinyarwanda	2.80	2.20	5.00	48	1172
pedi	2.18	1.94	4.13	33	1121
sesotho	2.56	2.11	4.68	28	1356
shona	2.81	2.18	4.99	41	1114
swahili	3.35	2.57	5.92	121	1377
tswana	2.09	3.56	5.65	51	1573
twi	0.58	0.13	0.71	4	339
xhosa	3.31	2.70	6.01	58	1799
yoruba	1.56	1.25	2.81	78	890
zulu	3.95	2.75	6.70	73	2331

Data column descriptions

• speaker_id [string]: speaker id for mapping to the audio file to the speaker
• audio_path [string]: path to the audio file
• transcription [string]: text transcription of the audio file
• translation [string]: text translation of the audio file
• domain [string]: domain (medical or non-medical)
• language [string]: language of the audio file
• gender [string]: gender of the speaker
• duration [float]: duration of the audio file in seconds