intronhealth/NigBench-MAMAI-Speech-QA

Voices for Smart Care

A Rural-First Multilingual Voice Dataset for Maternal Health in Nigeria

Voices for Smart Care is a multilingual speech dataset containing real-world maternal and reproductive health questions collected from women across Nigeria. The dataset was created to support the development and evaluation of Automatic Speech Recognition (ASR) and Large Language Models (LLMs) for low-resource African languages in healthcare settings.

Unlike generic speech datasets, Voices for Smart Care focuses on natural health-related conversations collected from the target population—particularly women living in rural and peri-urban communities—capturing realistic accents, dialects, code-switching, spontaneous speech, and environmental noise.

This release contains 300.17 hours across 10,081 recordings distributed across seven Nigerian languages: Hausa, Igbo, Nupe, Yoruba, Nigerian Pidgin, Fulfulde, and Kanuri.

---

Dataset Summary

The dataset contains audio recordings together with manually verified transcriptions and speaker/annotation metadata.

Features

Field	Type	Description
audio	audio	Speech recording (16 kHz mono)
sample_id	string	Per-sample id
speaker_id	string	Speaker id
language	string	Language identifier
split	string	train / test
transcriber_id	string	Transcriber id
admin_id	string	Recording-admin id
prompt_text	string	Clinical/health scenario prompt
audio_path	string	Source URL of the clip
audio_duration	float	Audio duration in seconds
transcript	string	Verbatim human transcription
level_speaker	string	Speaker annotation level
level_transcriber	string	Transcriber annotation level
neg_percent_speaker	string	Speaker negative-feedback rate
neg_percent_transcriber	string	Transcriber negative-feedback rate

The train/test split is speaker-disjoint (no speaker appears in both).

---

Languages

The dataset covers seven Nigerian languages.

Duration distribution (hours)

	language	test	train	Total
1	Fulfulde	2.41	20.88	23.29
2	Hausa	10.79	89.22	100.01
3	Igbo	3.88	37.45	41.33
4	Kanuri	1.45	12.78	14.23
5	Nupe	4.19	37.12	41.31
6	Pidgin	3.79	34.89	38.68
7	Yoruba	3.68	37.64	41.32
	Total	30.19	269.98	300.17

Sample distribution

	language	test	train	Total
1	Fulfulde	85 (10.4%)	731 (89.6%)	816
2	Hausa	350 (10.9%)	2861 (89.1%)	3211
3	Igbo	123 (10.1%)	1092 (89.9%)	1215
4	Kanuri	53 (10.8%)	439 (89.2%)	492
5	Nupe	126 (10.2%)	1105 (89.8%)	1231
6	Pidgin	146 (10.1%)	1304 (89.9%)	1450
7	Yoruba	155 (9.3%)	1511 (90.7%)	1666
	Total	1038 (10.3%)	9043 (89.7%)	10081

---

Data Collection

The Smart Care initiative aims to enable women across Nigeria to access maternal and reproductive healthcare information using voice interfaces in their native languages. Data collection prioritized rural and peri-urban communities, women with recent maternal healthcare experience, natural spoken health questions, and realistic recording environments. Rather than reading scripted prompts, participants described authentic maternal health concerns in their own words.

Each recording underwent manual transcription, language verification, and transcription quality review.

---

Intended Uses

• Automatic Speech Recognition (ASR)
• Speech foundation models & domain adaptation
• Healthcare NLP and multilingual LLM/RAG systems
• Benchmarking multilingual healthcare speech systems

---

Example

from datasets import load_dataset

# load a single language config, choose a split
ds = load_dataset("intronhealth/NigBench-MAMAI-Speech-QA", "Yoruba", split="test")

sample = ds[0]
print(sample["audio"])          # 16 kHz mono audio
print(sample["transcript"])     # human transcription
print(sample["language"], sample["split"])

---

Limitations

• Focuses on maternal and reproductive health; not representative of general conversational speech.
• Language coverage is intentionally imbalanced across the seven languages.
• Recordings include natural environmental noise, accents, and spontaneous speech.

Ethical Considerations

Data collection was conducted with informed participant consent. Personally identifying information was removed and speaker/transcriber/admin ids are anonymized. Intended solely for research and development of equitable healthcare technologies.

License

Released under the CC BY 4.0 license.