madoss/faso-speech

Faso Speech is a speech dataset built for Burkina Faso languages, starting with paired audio and text from Moore Burkina resources.

The current target languages are:

• Moore
• Dioula
• Fulfulde

French rows are preserved when they appear in the processed source material, mostly as translation or supporting content.

Dataset Summary

The dataset contains short audio chunks paired with transcripts. Source pages were archived separately, then processed into timed candidate chunks for ASR training and review.

The first source family is Moore Burkina. Many records embed or link IPS app-builder pages hosted on media.ipsapps.org; those pages expose timing arrays, text blocks, audio URLs, and next-page links used during extraction.

Current Processed Totals

These totals describe the processed dataset metadata used for this release.

Language	Rows	Duration	Rows shorter than 0.72s
moore	5,746	05:41:41.870	88
fulfulde	2,260	02:41:47.140	103
dioula	1,073	00:49:45.910	26
french	212	00:12:23.400	4

Configurations

Use one configuration per primary language:

• moore
• dioula
• fulfulde

Each configuration has deterministic train and validation splits. The split ratio is 90% train and 10% validation.

Data Fields

The dataset uses a compact ASR-oriented schema:

Column	Description
audio	Embedded Hugging Face Audio(decode=False) value with bytes and path
text	Transcript
language	Language label
duration	Chunk duration from metadata
chunk_id	Stable chunk identifier
record_id	Source record identifier
content_type	Source content type

Usage

from datasets import Audio, load_dataset

ds = load_dataset("madoss/faso-speech", "moore")
ds = ds.cast_column("audio", Audio(decode=False))

print(ds["train"][0]["text"])
print(ds["train"][0]["audio"])

Audio Preparation

The dataset is distributed as embedded Hugging Face Audio(decode=False) examples. When segment metadata was available during preparation, only detected speech ranges were embedded, with small padding around speech to avoid clipping words. Rows without segment metadata keep their original chunk audio.

Preparation used the following default padding and merge behavior:

Option	Default
--segment-start-padding	0.15
--segment-end-padding	0.25
--music-start-padding	0.05
--music-end-padding	0.05
--max-intra-segment-gap	0.50
--max-music-gap	0.05

Source Data And Provenance

The source archive preserves raw source artifacts separately from processed training examples, including source URLs, app-builder HTML, and optional wrapper HTML, audio URLs, downloaded audio, extracted text blocks, timing metadata, language labels, content type, and metadata.

This separation keeps the processed dataset traceable to the original source material.