metadata
language:
- af
- ak
- am
- ar
- fr
- ff
- gaa
- ha
- ig
- rw
- lg
- nso
- st
- sn
- sw
- tn
- tw
- xh
- yo
- zu
AfriVox-Transcribe: An African benchmark dataset for Automatic Speech Recognition
Project Overview
This project creates a benchmark dataset for evaluating Automatic Speech recognition models on African languages. This benchmark dataset combines ASR test sets from multiple domain and sources for 20 African languages. See the tables below for source and language details.
License
The dataset incorporates various licenses inherited from its different sources. Each source’s license governs how you may use, modify, and redistribute its respective data. Please review the linked license for each source to ensure your use complies with its terms.
- ALFFA: MIT
- Ashesi FISD: CC BY 4.0
- NCHLT: CC BY 3.0
- BibleTTS: CC-BY-SA
- common voice: CC0 1.0 Universal
- fleurs: CC BY 4.0
- iwslt_lrst: IWSLT License
- naijavoices: CC BY-NC-SA 4.0
- openslr: CC BY-SA 3.0
- intron-MT: CC BY-NC-SA 4.0
Key Objectives / Goals
Usage Instructions
Accessing the Dataset: The dataset can be accessed through Hugging Face:
from datasets import load_dataset
afrivox_transcribe = load_dataset("intronhealth/afrivox-transcribe")
Afrivox stats
- Dataset size = 41,733
- Total number of languages = 20
- Total number of sources = 9
- Total number of hours = 79.47
Source distribution (hours)
| Source | Duration (hours |
|---|---|
| ALFFA | 0.95 |
| Ashesi-Org_Financial-Inclusion-Speech-Dataset (FISD) | 0.05 |
| NCHLT | 12.75 |
| bible_tts | 0.47 |
| common_voice (CV) | 16.89 |
| fleurs | 14.44 |
| intron-MT | 30.63 |
| naijavoices | 2.98 |
| openslr | 0.31 |
Duration per Language:
| Language | Duration (hours) |
|---|---|
| Afrikaans | 4.64 |
| Akan | 1.47 |
| Amharic | 1.23 |
| Arabic | 4.94 |
| French | 0.76 |
| Fulani | 0.32 |
| Ga | 0.03 |
| Hausa | 3.43 |
| Igbo | 2.88 |
| Kinyarwanda | 5.53 |
| Luganda | 3.54 |
| Pedi | 6.72 |
| Sesotho | 5.73 |
| Shona | 6.37 |
| Swahili | 5.45 |
| Tswana | 6.17 |
| Twi | 0.75 |
| Xhosa | 6.68 |
| Yoruba | 5.77 |
| Zulu | 7.08 |
Language Duration (hours) per Source
| language | ALFFA | FISD | NCHLT | bible_tts | CV | fleurs | intron-MT | naijavoices | openslr |
|---|---|---|---|---|---|---|---|---|---|
| afrikaans | - | - | 2.66 | - | 0.11 | 0.07 | 1.79 | - | - |
| akan | - | 0.01 | - | - | - | - | 1.47 | - | - |
| amharic | 0.73 | - | - | - | 0.35 | 0.15 | - | - | - |
| arabic | - | - | - | - | 3.63 | 0.04 | 1.27 | - | - |
| french | - | - | - | - | - | - | 0.44 | - | 0.31 |
| fulani | - | - | - | - | - | 0.32 | - | - | - |
| ga | - | 0.03 | - | - | - | - | - | - | - |
| hausa | - | - | - | 0.13 | 0.77 | 0.32 | 1.23 | 0.99 | - |
| igbo | - | - | - | - | 0.00 | 0.42 | 1.46 | 1.00 | - |
| kinyarwanda | - | - | - | - | 3.99 | - | 1.53 | - | - |
| luganda | - | - | - | - | 3.46 | 0.08 | - | - | - |
| pedi | - | - | 1.85 | - | - | 2.04 | 2.82 | - | - |
| sesotho | - | - | 2.52 | - | - | - | 3.20 | - | - |
| shona | - | - | - | - | - | 3.82 | 2.55 | - | - |
| swahili | 0.22 | - | - | - | 3.34 | 0.34 | 1.55 | - | - |
| tswana | - | - | 2.39 | - | - | - | 3.78 | - | - |
| twi | - | 0.01 | - | 0.22 | 0.01 | - | 0.50 | - | - |
| xhosa | - | - | 1.52 | - | - | 2.20 | 2.97 | - | - |
| yoruba | - | - | - | 0.13 | 1.22 | 2.69 | 0.74 | 1.00 | - |
| zulu | - | - | 1.80 | - | - | 1.96 | 3.32 | - | - |
Data column descriptions
- speaker_id [string]: speaker id for mapping to the audio file to the speaker
- file_name [string]: path to the audio file
- transcription [string]: text transcription of the audio file
- source [string]: source of the audio file
- language [string]: language of the audio file
- gender [string]: gender of the speaker
- duration [float]: duration of the audio file in seconds