The Thiomi Dataset: A Large-Scale Multimodal Corpus for Low-Resource African Languages

Hillary Mutisya; John Mugane; Gavin Nyamboga; Brian Chege; Maryruth Gathoni

cs.CLMar 31, 2026

The Thiomi Dataset: A Large-Scale Multimodal Corpus for Low-Resource African Languages

Authors: Hillary Mutisya, John Mugane, Gavin Nyamboga, Brian Chege, Maryruth Gathoni

Organizations: Thiomi NLP · Harvard University

Abstract

We present the Thiomi Dataset, a large-scale multimodal corpus spanning ten African languages across four language families: Swahili, Kikuyu, Kamba, Kimeru, Luo, Maasai, Kipsigis, Somali (East Africa); Wolof (West Africa); and Fulani (West/Central Africa). The dataset contains over 601,000 approved sentence-level text annotations and over 385,000 audio recordings, collected through a dedicated community data collection platform involving over 100 contributors. To validate the dataset's utility, we train and evaluate ASR, MT, and TTS models, establishing baselines across all languages. Our best ASR system achieves 3.24% WER on Swahili (Common Voice), reducing prior academic SOTA from 8.3% to 3.24% (5.1 percentage point absolute, 61% relative reduction), and 4.3% WER on Somali. The dataset will be published on HuggingFace. We describe the collection platform, quality assurance workflows, and baseline experiments, and discuss implications for African language technology infrastructure.

Figures & tables

	Langs	Hours	Modal.	Baselines	Region
Common Voice	100+	20K+	Audio	ASR	Global
FLEURS	102	12	Audio	ASR	Global
MMS	1100+	44K	Audio	ASR,TTS	Global
WAXAL	24	1485	Audio	—	W/C Africa
Thiomi	10	1500	Both	ASR,MT,TTS	E Africa

Table 1: Comparison with existing multilingual speech datasets. The Thiomi Dataset is the only resource providing integrated text+audio with trained baselines for East African languages.

Language	Family	Spk (M)	Script	Location
Swahili	Bantu G42	200	Latin	E Africa
Kikuyu	Bantu E51	8.1	Latin	Kenya
Kamba	Bantu E55	4.6	Latin	Kenya
Kimeru	Bantu E54	3.5	Latin	Kenya
Luo	Nilotic	6.0	Latin	Kenya/Uganda
Maasai	E. Nilotic	1.5	Latin	Kenya/Tanzania

Table 2: Languages in the Thiomi Dataset. The ten languages span four genealogical families and represent over 300 million speakers. Guthrie zone codes (e.g., G42) are standard Bantu language classification identifiers based on geographic and linguistic groupings.

Panel of Figure 1 — Figure 1: Two-pipeline data collection workflow. Text-first (left) produces aligned English–target pairs with recorded audio. Audio-first (right) captures naturalistic speech with community transcription. Both feed into the same 4-stage QA pipeline.

Lang	Sentence pair
eng	The doctor advised her to rest for two days.
swh	Daktari alimwambia apumzike kwa siku mbili.
eng	We planted maize in the garden last week.
kik	Twahandire mbembe mũgũnda kiumia kĩrĩa kĩathĩire.
eng	The children are playing outside.
som	Caruurtu waxay ku ciyaaraysaa dibadda.

Table 3: Sample English–target sentence pairs from the Thiomi Dataset (health, agriculture, and daily life domains).

Domain	Coverage
Daily life and household	12%
Health and medical	11%
Agriculture and environment	11%
Education and school	10%
Government and civic	10%
Business and commerce	10%

Table 4: Domain distribution in the Thiomi Dataset text corpus.

Panel of Figure 2 — Figure 2: Approved text sentences per language. Six languages exceed 64,000 approved sentences. Kipsigis, Maasai, and Wolof are in earlier collection stages. Total: 601,409 approved sentences across 9 languages.

Language	Translated	Approved	Recordings	Appr.%
Kikuyu	103,669	102,294	71,334	98.7%
Luo	103,693	101,367	87,804	97.8%
Somali	103,086	103,086	97,537	100.0%
Kimeru	99,582	94,933	74,586	95.3%
Kamba	103,554	89,504	54,469	86.4%
Fulani	70,143	64,567	0	92.1%

Table 5: Per-language collection statistics from the Thiomi platform (as of March 2026). Translated = sentences completed by community translators; Approved = sentences passing peer moderation; Recordings = audio clips recorded; Appr.% = text approval rate. Audio recording is ongoing for Fulani, Maasai, and Kipsigis.

Split	Proportion	Text	Audio
Train	80%	$\sim$ 481,000	$\sim$ 309,000
Development	10%	$\sim$ 60,000	$\sim$ 38,000
Test	10%	$\sim$ 60,000	$\sim$ 39,000

Table 6: Standard train/dev/test splits. Test set sentences were collected independently from training sentences by a disjoint set of contributors. Contributor disjointness is enforced by the platform: contributors are assigned to either the training or evaluation pool at registration, and the pools do not overlap.

Language	WER (CV)	Prior SOTA	Improvement
Swahili	3.24%	8.3% (XLS-R)	61% reduction
Somali	4.3%	n/a	—
Kikuyu	5.5%	n/a	—
Kamba	6.2%	n/a	—
Kimeru	6.8%	n/a	—
Luo	7.1%	n/a	—

Table 7: ASR Word Error Rate (WER) results. Swahili (3.24% WER) surpasses the prior academic state of the art (XLS-R finetuned, 8.3%) by 61% relative improvement, using continued pretraining on pseudo-labeled unlabeled audio ( Mutisya and Mugane, 2026 ) . All other languages report the first published results.

Language pair	Dir.	BLEU	Train pairs
Somali–English	som $\to$ eng	64.2	60,002
	eng $\to$ som	53.2
Swahili–English	swh $\to$ eng	55.8	80,214
	eng $\to$ swh	48.3
Luo–English	luo $\to$ eng	44.64	60,127
	eng $\to$ luo	34.95

Table 8: Machine translation BLEU scores for the 6 language pairs supported by NLLB-200. Maasai, Kipsigis, Wolof, and Fulani are not in NLLB’s language inventory and require from-scratch MT training (planned for future work). Into-English consistently outperforms from-English.

Language	MOS	Natural.	Intellig.
Swahili	4.12	4.21	4.03
Kikuyu	3.87	3.91	3.83
Kamba	3.84	3.88	3.80
Somali	3.76	3.82	3.70
Luo	3.71	3.75	3.67
Kimeru	3.69	3.72	3.66

Table 9: TTS Mean Opinion Score (MOS, 5-point scale). Each language was evaluated by 5 native speaker evaluators on 50 synthesized utterances. Evaluators rated naturalness and intelligibility independently on a 1–5 Likert scale. Evaluators were not involved in data collection.

Task	Architecture	Languages
ASR	Wav2Vec2-BERT (607M)	6
MT	NLLB-600M fine-tuned	5 pairs (bidir.)
TTS	VITS ( $\sim$ 30M)	4

Table 10: Baseline experiments conducted to validate dataset quality. Models trained for evaluation purposes; model release is planned for future work.

African Language Processing

See all 34 papers

From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa
Jun 20, 2026Mahounan Pericles Adjovi, Victor Olufemi, Roald Eiselen +1Automatic Speech Recognition Low-Resource Language Processing
Carnegie Mellon University Africa, Kigali, Rwanda · North-West University, South Africa
AFRILANGTUTOR: Advancing Language Tutoring and Culture Education in Low-Resource Languages with Large Language Models
Apr 22, 2026Tadesse Destaw Belay, Shahriar Kabir Nahin, Israel Abebe Azime +6Multilingual Language Models Intelligent Tutoring Systems
Instituto Politécnico Nacional, Mexico · University of South Florida, FL, USA · Saarland University, Germany +2
AfriVoices-KE: A Multilingual Speech Dataset for Kenyan Languages
Apr 9, 2026Lilian Wanzare, Cynthia Amol, Ezekiel Maina +14Low-Resource Language Processing Low-Resource Speech Recognition
Maseno University, Kenya · United States International University, Kenya · Dedan Kimathi University of Technology, Kenya +4

Jun 20, 2026J/K move · Enter open · S save

From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa

Mahounan Pericles Adjovi, Victor Olufemi, Roald Eiselen, Prasenjit Mitra

Carnegie Mellon University Africa, Kigali, Rwanda · North-West University, South Africa

Automatic Speech Recognition Low-Resource Language Processing Low-Resource Speech Recognition Multilingual ASR African Language Processing

Open paper AskPDF arXiv

Low-resource African languages lack text corpora needed for language model training. We investigate whether ASR pipelines can extend text resources for two typologically distinct West African languages: Fongbe (tonal, diacritic-rich) and Hausa (non-tonal). We fine-tune MMS-300M on a curated 12.3-hour Fongbe dataset, achieving 9.48% WER on the ALFFA benchmark - a 78% relative reduction from the prior 44.04% baseline - while preserving tonal diacritics critical to the language. For Hausa, we apply an existing fine-tuned Whisper-Small model. We catalog 1,553 YouTube videos (236 hours) and process a subset of 424 videos (45.49 hours) selected to balance domain diversity with available computational resources, producing 6,770 transcribed segments. Human evaluation on 50 randomly sampled segments per language shows mean quality scores of 57.4/100 for Hausa and 36.5/100 for Fongbe, indicating that while Hausa transcriptions approach acceptable quality for corpus construction, Fongbe transcriptions require post-processing or improved models for production use. We release the curated dataset, fine-tuned model, transcribed corpus, and full video catalog following platform terms and ethical guidelines.

Abstract

African Language Processing

Explore similar work

From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa

AFRILANGTUTOR: Advancing Language Tutoring and Culture Education in Low-Resource Languages with Large Language Models

AfriVoices-KE: A Multilingual Speech Dataset for Kenyan Languages