Datasets:
AfriVox-v2: A Domain-Verticalized Benchmark for In-the-Wild African Speech Recognition
AfriVox-v2 is a comprehensive multilingual speech recognition benchmark designed to evaluate ASR systems under realistic African deployment conditions. It covers 24 African languages across 10 application domains, with a strong emphasis on spontaneous, unscripted "in the wild" audio.
Dataset Summary
Most existing ASR benchmarks for African languages rely on scripted, read speech — a setting that fails to capture the complexity of real-world voice interfaces. AfriVox-v2 addresses this by aggregating conversational speech corpora, in-the-wild datasets, and providing granular domain-level annotations across sectors such as health, finance, agriculture, and government.
Languages
AfriVox-v2 covers 24 African languages spanning multiple language families:
| Language | Code | Utterances |
|---|---|---|
| Nigerian Pidgin | pcm |
12,546 |
| Kinyarwanda | kin |
12,493 |
| Yoruba | yor |
10,911 |
| Oromo | orm |
7,475 |
| Igbo | ibo |
7,229 |
| Amharic | amh |
6,317 |
| Sesotho | sot |
6,317 |
| Swahili | swa |
6,074 |
| Xhosa | xho |
6,041 |
| Zulu | zul |
5,586 |
| Tswana | tsn |
5,176 |
| Wolof | wol |
4,670 |
| Fulani | ful |
4,098 |
| Hausa | hau |
3,691 |
| Arabic | ara |
3,413 |
| Afrikaans | afr |
3,355 |
| Luganda | lug |
2,768 |
| Bemba | bem |
2,768 |
| Pedi | nso |
2,455 |
| Akan | aka |
1,973 |
| Shona | sna |
1,185 |
| French | fra |
379 |
| Twi | twi |
378 |
| Ga | gaa |
16 |
Data Sources
AfriVox-v2 aggregates three speech corpora:
Africa Next Voices (AFN)
A multi-domain decentralized African conversational speech corpus funded by the Gates Foundation. Regional grantees created and hosted data separately. For subsets with explicit test splits, those were used directly; for others, an 80/10/10 train/dev/test split was created, stratified by speaker to prevent leakage.
- Languages: 15
- Hours: ~100
- Type: Conversational
Waxal
A multilingual African language speech corpus funded by Google, drawn from the full Waxal corpus.
- Languages: 6
- Hours: ~69.5
- Type: Conversational
AfriVox (v1)
The original AfriVox benchmark (Awobade et al., EACL 2026), which aggregates multiple open-source corpora (NCHLT, Common Voice, FLEURS, OpenSLR, BibleTTS, NaijaVoices, FISD) alongside the novel AfriVox-Medical dataset — health-focused read speech across 20 African languages transcribed and translated by native speakers. Covers both read and medical domain speech.
- Languages: 20
- Hours: ~81.5 (multilingual subset)
- Type: Read speech / Medical domain
All audio has been resampled to 16kHz.
Domain Taxonomy
AfriVox-v2 introduces a unified 10-domain taxonomy for fine-grained evaluation of real-world application performance:
| Domain | Sentences | Duration (hrs) |
|---|---|---|
| General | 20,117 | 29.67 |
| Health | 7,321 | 22.70 |
| Culture & Society | 4,434 | 7.58 |
| Government | 4,323 | 8.76 |
| Finance | 1,882 | 3.31 |
| Education | 1,793 | 3.74 |
| Agriculture | 947 | 1.38 |
| Transportation | 911 | 1.82 |
| Sports & Hobbies | 620 | 1.47 |
| Telecommunications | 470 | 1.31 |
Domain labels were generated using an LLM-assisted multilabel tagging pipeline and validated by human annotators on a random sample of ~50 utterances per language per tag, with priority given to the 6 highest-utterance languages.
In addition to the 10 standard domains, two targeted subsets are provided:
- Named Entities (EWER): Utterances containing at least one named entity
- Numbers (NWER): Utterances containing at least one numerical expression
Dataset Structure
Configs / Subsets
The dataset is organized into three configs corresponding to the source corpora:
| Config | Source | Description |
|---|---|---|
WAXAL |
Waxal corpus | Conversational speech across 6 African languages |
AFN |
Africa Next Voices | Multi-domain conversational speech across 15 languages |
Afrivox |
AfriVox v1 | Read + medical domain speech across 20 languages |
Fields
Each example contains the following fields:
| Field | Type | Description |
|---|---|---|
audio |
Audio (16kHz) |
Audio signal |
language |
string |
ISO 639-3 language code |
reference |
string |
Ground-truth transcript |
domains |
List[string] |
Domain tag(s) assigned to the utterance |
has_named_entities |
bool |
Whether the utterance contains named entities |
has_numbers |
bool |
Whether the utterance contains numerical expressions |
named_entities |
List[string] |
Named entities present in the transcript |
source |
string |
Source corpus (WAXAL, AFN, or Afrivox) |
Loading the Dataset
from datasets import load_dataset
# Load a specific config
ds = load_dataset("intronhealth/afrivox-v2", "AFN", split="test")
# Load all configs
from datasets import concatenate_datasets
configs = ["WAXAL", "AFN", "Afrivox"]
ds = concatenate_datasets([
load_dataset("intronhealth/afrivox-v2", cfg, split="test")
for cfg in configs
])
Quality Assurance
AfriVox (v1): Transcriptions and translations were produced by college-educated bilingual native speakers. A two-stage review pipeline was applied: primary transcription/translation by native speakers, followed by independent meta-review where graduate-level annotators validated 10–20% of each contributor's output. Contributors achieving less than 80% accuracy were excluded. Full details in the AfriVox paper.
AFN and Waxal: Samples with anomalously high WER from a reference model were flagged, then manually reviewed against transcripts to remove bad audio and mismatched pairs. Approximately 90% of samples were retained, confirming overall corpus quality.
Intended Use
AfriVox-v2 is intended for:
- Benchmarking ASR models on African languages under real-world conditions
- Domain-specific evaluation of speech systems in sectors like health, finance, and agriculture
- Error analysis on numerically and entity-dense transcription tasks
- Research into low-resource and underrepresented language ASR
It is not intended for model training (it is a test/evaluation benchmark).
Citation
If you use AfriVox-v2 in your research, please cite:
@article{awobade2026afrivox,
title = {AfriVox-v2: A Domain-Verticalized Benchmark for In-the-Wild African Speech Recognition},
author = {Awobade, Busayo and Ashungafac, Gabrial Zencha and Olatunji, Tobi},
institution = {Intron Health},
year = {2026}
}
Contact
For questions or feedback, reach out to the Intron Health research team at research@intron.io.
- Downloads last month
- 677