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), Machine Translation (MT), 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 150.66 hours across 5,323 recordings, and represents a subset of the planned 600-hour corpus (100 hours each for Hausa, Igbo, Yoruba, Fulfulde, and Nigerian Pidgin; 50 hours each for Nupe and Kanuri). Language coverage is intentionally imbalanced at this stage to reflect data available at time of release.

Dataset Summary

The dataset contains audio recordings together with manually verified transcriptions in seven Nigerian languages.

Features

Field	Type	Description
`audio`	`audio`	Speech recording
`text`	`string`	Human transcription
`language`	`string`	Language identifier
`duration`	`float`	audio duration in seconds

The dataset is intended for:

Automatic Speech Recognition (ASR)
Speech model evaluation
Speech translation research
Domain adaptation
Healthcare NLP
Low-resource language modeling

Languages

The dataset covers seven Nigerian languages.

Language	Hours	Recordings
Hausa	51.85	1,559
Yoruba	51.84	2,074
Nigerian Pidgin	25.95	1,026
Nupe	8.45	257
Igbo	4.40	117
Kanuri	4.37	154
Fulfulde	3.80	136
Total	150.66	5,323

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 communities
Peri-urban communities
Women with recent maternal healthcare experience
Natural spoken health questions
Realistic recording environments

Speech was collected through multiple channels including:

Community health centres
Rural clinics
WhatsApp voice recordings
Assisted recordings by trained field researchers

Rather than reading scripted prompts, participants were encouraged to describe authentic maternal health concerns in their own words.

Collection Methodology

The dataset was designed using an error-driven data collection strategy.

1. Error Analysis

Existing multilingual maternal-health speech datasets were first used to benchmark current ASR systems and identify weaknesses such as:

medical terminology
symptoms
medication names
pregnancy timelines
numerical expressions
dialect variation
environmental robustness

These findings informed the design of subsequent data collection.

2. Population-Matched Recruitment

Recruitment focused primarily on rural and peri-urban women to better reflect the intended users of Smart Care voice assistants.

3. Natural Speech

Participants described maternal and reproductive health situations using their own language rather than reading scripted sentences.

Typical queries include:

pregnancy symptoms
antenatal care
newborn care
breastfeeding
nutrition
malaria
fever
labour
postpartum recovery
medication questions

4. Quality Assurance

Each recording underwent:

manual transcription
language verification
transcription quality review

Intended Uses

The dataset is intended for research on:

Automatic Speech Recognition
Speech Foundation Models
Domain Adaptation
Healthcare AI
Speech Translation
Machine Translation
Large Language Models
Retrieval-Augmented Generation
Benchmarking multilingual healthcare systems

Example

from datasets import load_dataset

dataset = load_dataset("intronhealth/Nig-Voice-Health-Bench")

sample = dataset["train"][0]

print(sample["audio"])
print(sample["text"])
print(sample["language"])

Acknowledgements

This dataset was developed as part of the Smart Care initiative to improve equitable access to maternal healthcare information through multilingual voice technologies.

We gratefully acknowledge the contributions of participating women, community health workers, healthcare facilities, field researchers, transcribers, and language experts who made this resource possible.

Limitations

The dataset focuses on maternal and reproductive health and should not be considered representative of general conversational speech.
Language coverage is intentionally imbalanced to reflect available data at the time of release.
Recordings include natural environmental noise, accents, and spontaneous speech, which may increase task difficulty but better reflect real-world deployment conditions.
This release is a partial (~150-hour) subset of the planned 600-hour corpus and should not be treated as the final or complete collection.

Ethical Considerations

Data collection was conducted with informed participant consent. Personally identifying information was removed during processing. The dataset is intended solely for research and development of equitable healthcare technologies and should not be used to infer sensitive personal attributes about speakers.

License

This dataset is released under the CC BY-NC-SA 4.0 license.

Downloads last month: 1

Total file size:

69.1 GB