African-Languages-Lab/gun-asr

Gun ASR Speech Dataset

A speech dataset in Gun (guw) for training and evaluating automatic speech recognition (ASR) models. It is part of an ASR project covering Fon, Gun and Baoulé. Each recording is a speaker describing an image prompt in their own words (semi-spontaneous speech), paired with its transcription and the image.

Dataset summary

Language	Gun (ISO 639-3 guw)
License	CC BY 4.0
Task	Automatic speech recognition
Speech type	Semi-spontaneous (image description)
Utterances	712
Transcribed	712 of 712
Speakers	6
Total audio	5.11 hours (5h 06m 22s)
Average utterance	25.8 s
Recording country	Benin

Speakers

Speakers	6
Gender	4 male, 2 female
Age range	5 25-34, 1 35-44

Speakers are identified only by an opaque speaker_id. Use it to keep each speaker in a single split when making train/test splits, so evaluation is on unseen voices.

Audio specification

Format	WAV (PCM)
Sample rate	48 kHz
Bit depth	16-bit
Channels	Mono
Signal-to-noise ratio	≥ 40 dB
Duration per file	10–30 s (average 25.8 s)

Domain distribution

Domain	Utterances	Hours	Share of total hours
Agriculture	283	2.11	41.3%
Finance	190	1.33	26.1%
Health	72	0.49	9.6%
Government	54	0.35	6.9%
Education	113	0.82	16.1%

Data collection and quality

Recordings were collected through a recording app. Speakers were shown an image from one of the domains above and asked to describe it in their own words. Quality measures have been put in place at each step:

• Every recording was reviewed by validators; only recordings with status approved are included.
• Recordings shorter than 10 seconds are excluded.
• Recordings must have a signal-to-noise ratio of at least 40 dB.
• Audio must be 48 kHz, 16-bit, mono WAV.
• Every row was checked to be in Gun (guw).
• Metadata and media files were checked for consistency, so every row points to its own audio and image.

Dataset structure

audio/<audio_id>.wav
images/<image>.jpg
metadata.csv   (UTF-8 with BOM)
metadata.json
README.md

metadata.csv and metadata.json hold the same rows in the same order. All data is in a single train split.

Data fields

Field	Type	Description
audio_id	string	Unique ID of the recording
audio	string / audio	Audio file name; on the Hugging Face Hub, the playable audio
audio_path	string	Path of the WAV file inside the dataset folder
transcription	string	Transcription of the speech
speaker_id	string	Opaque speaker identifier (no names are included)
speaker_gender	string	Speaker gender
speaker_age_range	string	Speaker age range, e.g. 25-34
speaker_country	string	Speaker country
duration	float	Length of the recording in seconds
source_language	string	ISO 639-3 language code
dataset_tags	string	Domain of the prompt (comma-separated if several)
prompt_type	string	Type of prompt; semi_spontaneous = describing an image in one's own words
image	string	ID of the image prompt
image_path	string	Path of the image prompt inside the dataset folder

Example

{
  "audio_id": "5882",
  "audio": "5882.wav",
  "audio_path": "audio/5882.wav",
  "transcription": "Mǐ mọn ákokǔɛ́ lẹ ǔnsọ ǔnsọ yɛ́tọn kokǔɛ́ wẹ yɛ́vɖo xíá mǐ mí mọn ɖo ákòkǔɛ́ lọ lɛ̧ ɖɛ́kpokpɛ́ mán biá àgloɛ́ lẹ kẹɖẹ wẹ yɛ́ tǐn to sọnǔ sónǔ mǐ to mímɔ̧ to ɛ́xɛ́ yǐn yẹ ɛ́xɛ́ lẹ mẹ fǐ",
  "speaker_id": "cabbf96f-4f68-454c-ab14-daf72f5c942a",
  "speaker_gender": "male",
  "speaker_age_range": "25-34",
  "speaker_country": "Benin",
  "duration": 24.9,
  "source_language": "guw",
  "dataset_tags": "Agriculture",
  "prompt_type": "semi_spontaneous",
  "image": "5882",
  "image_path": "images/5882.jpg"
}

Usage

From the Hugging Face Hub:

from datasets import load_dataset

ds = load_dataset("all-lab/gun-asr", split="train")
print(ds[0]["transcription"])

From the delivered folder:

import pandas as pd

df = pd.read_csv("metadata.csv", encoding="utf-8-sig")

License

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