Data-Science-Nigeria/voice-of-care-health-dataset

Voice of Care AI for Global Health Benchmark Dataset

Overview

This dataset contains spoken Hausa Health datasets with rich annotations covering emotion, intent, speaker demographics, and dialect variation, intended for speech and NLP research.

Dataset Summary

🛠️ Dataset Structure and Features

Summary Statistics

• —Count of Participants: 50
• —Number of Dialects: 4
• —Male Speakers: 25
• —Female Speakers: 25
• —Number of Unique Audio Clips: 30,000
• —Number of Intents: 13
• —Female Audio Clips: 15,000
• —Male Audio Clips: 15,000
• —Total Hours: 52 hrs

User ID Convention

Format: {ORG}{GENDER}{DIALECT}{AGEGROUP}_{NUMBER}

Example: EHAFKN0001

Where:

• —EHA = Organization
• —Gender
• —Female = F
• —Male = M
• —Dialect
• —Kananci = KN
• —Katsinanci = KT
• —Sakkwatanci = SK
• —Zazzaganci = ZZ
• —Age Group
• —15–29 = 0
• —30–45 = 1
• —45+ = 2 Number = Sequential participant number padded to 3 digits (e.g., 001, 002, 003)

Data Collection

The prompts were provided by the eHealth team and subsequently translated into Hausa. Each prompt was carefully classified according to predefined categories, including emotion categories, micro-intents, speaker type, patient statements, patient questions, and caregiver questions. After the classification process, the prompts were converted into speech using the EqualyzAI data collection platforms (Equalyz Crowd and Equalyz Gram). The generated voice recordings were then reviewed and validated by the EHA team through the Equalyz Gram platform to ensure quality, accuracy, and consistency with the intended content.

💻 Loading the Dataset (Hugging Face)

The dataset is configured for easy loading of the dataset.

Recommended Environment

pip install --upgrade datasets[audio]
pip install --upgrade ffmpeg ffmpeg-python

Standard Load

from datasets import load_dataset

# Load the full train
ds_train = load_dataset("Data-Science-Nigeria/voice-of-care-health-dataset", "default", split="train")

# Load a specific split (e.g., development)
ds_test = load_dataset("Data-Science-Nigeria/voice-of-care-health-dataset", "default", split="test")

Streaming Mode (for memory efficiency)

from datasets import load_dataset

# Load the dev_test split in streaming mode
ds_stream = load_dataset(
    "Data-Science-Nigeria/voice-of-care-health-dataset", 
    "default", 
    split="test", 
    streaming=True
)

Benchmarking

Benchmark Results

The Hausa Health Spoken Dialogue Corpus was benchmarked across two core capabilities:

1. 1.Intent Classification — measuring the ability of machine learning models to identify the intent expressed in patient and caregiver statements.
2. 2.Automatic Speech Recognition (ASR) — evaluating the quality and accuracy of speech transcription models on Hausa healthcare conversations.

Intent Classification Benchmark

A machine learning model was trained and evaluated to classify the intent behind patient and caregiver statements using annotated transcripts and corresponding intent labels.

The model's objective is to automatically identify the purpose of a speaker's message, such as reporting symptoms, requesting information, discussing clinical assessments, expressing medication concerns, or requesting follow-up support. This enables faster triage, structured analysis of conversational data, and development of healthcare-focused AI applications.

On a held-out test set, the model achieved an overall accuracy of 96.1%. Performance was strong across the most common intent categories, while less frequent categories showed lower reliability due to limited examples. Additional data collection for underrepresented intents is expected to further improve model performance.

Automatic Speech Recognition (ASR) Benchmark

The dataset was evaluated using several speech recognition models to measure transcription quality for Hausa healthcare conversations.

Evaluation metrics:

• —Word Error Rate (WER): Measures the percentage of incorrectly transcribed words. Lower values indicate better transcription performance.
• —Character Error Rate (CER): Measures the percentage of incorrectly transcribed characters and provides a more fine-grained measure for languages with complex morphology, such as Hausa.

Models evaluated:

• —Whisper Large-v3
• —XLSR-53
• —MMS-1B-fl102

🚀 Intended Use & Applications

This dataset is specifically designed for the following purposes:

• —Automatic Speech Recognition (ASR) Training
• —NLP for low-resource African languages
• —Cross-lingual learning and transfer learning research
• —Evaluation of multilingual ASR systems
• —Linguistic research and accent/dialect modeling

⛔ Use Restrictions & Limitations

To protect speaker privacy and prevent potential voice misuse, the following uses are strictly enforced:

Strictly Prohibited Uses 🛑

• —Voice cloning or adaptation (Text-to-Speech/TTS)
• —Voice biometrics, speaker identification, or mimicry
• —Surveillance, profiling, or any commercial application that relies on identifying or imitating individuals

Limitations ⚠️

• —Not suitable for biometric or forensic use

⚖️ License and Citation

License

This dataset is released under the Creative Commons Attribution 4.0 (CC BY 4.0) license.

📧 Contact & Support

For questions, issues, or collaboration inquiries, please open an issue on the repository or contact the maintainers directly.

Maintainer: This dataset was commissioned by eHealth Africa Clinics (EHA) and developed in partnership with Data Science Nigeria / EqualyzAI as part of the Voices of Care AI for Global Health Benchmarking Initiative, funded by The Agency Fund.