Afri-names: Read Speech Dataset of Numbers and African Named Entities

This work is licensed under a
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Overview

Afri-names is a curated African-accented read speech dataset comprising 6,307 single-speaker audio samples, totaling 8.92 hours of speech data. Each sample is densely populated with numbers or African named entities or voice commands (with African named entities), making it ideal for ASR (Automatic Speech Recognition) and Named Entity Recognition (NER) tasks. The recordings span 12 distinct African accents across 4 countries.

Dataset Summary

	Numbers	Names	Commands	Combined
Num. of Samples	914	3,121	2,272	6,307
Avg. Duration per Sample (seconds)	15.43	2.52	4.46	5.64
Total Duration (hrs)	3.92	2.18	2.81	8.92
Countries Represented	2	3	3	4
Distinct Accents	4	6	5	12
Age Range (yrs)	19–55	19–55	19-40	19–55

Use Cases

• Evaluation of ASR systems on African-accented English
• Testing Named Entity Recognition (NER) pipelines on spoken data

Dataset Structure

Each row in the dataset includes:

• file_name: Path to the .wav audio file
• transcript: The provided transcript corresponding to the audio
• domain: Either "numbers", "names" or "commands"
• accent: Primary accent of the speaker
• country: Country code of the speaker (e.g., NG, ZA)
• age_group: Age group of the speaker
• duration: Duration of the recording in seconds

All recordings are single-speaker, recorded at 16kHz in mono-channel format.

Accents and Countries

• Accents: Ijaw, Yoruba, Igbo, Shona, Akan, Swahili, Kikuyu, Kamba, Hausa, Zulu, Tswana
• Countries: Nigeria (NG), South Africa (ZA), Ghana (GH), Kenya (KE)

Data Collection and Processing

• Collection Method: Text prompts (names or numbers) were provided to contributors, who recorded the corresponding audios in varied, but generally quiet, acoustic environments.
• Annotation: As texts were predefined, all transcripts are clean and verified. No additional timestamp-level annotation is provided.

Data Split

This release is a test-only dataset and does not include predefined train/dev splits.

How to Load the Dataset

from datasets import load_dataset

dataset = load_dataset("intronhealth/afri-names")