madoss/faso-speech-plus

Faso Speech Plus

Faso Speech Plus extends madoss/faso-speech by concatenating it with other already-published ASR datasets covering the same target languages (Mooré, Jula, Fulfuldé), plus related language varieties grouped by macrolanguage.

Built with scripts/build_faso_speech_plus.py. See that script's MANIFEST for the authoritative, per-source list of dataset ids, column mappings, target config, license, and attribution.

Configs and dialect handling

Config	Language group	Row-level language values
moore	Mooré	mos (Mooré)
bambara_jula	Bambara and Jula	dyu (Jula/Dioula), und when the merged source does not identify the variety
fula	Fula	ffm, fuc, fue, fuf, fuh, fui, fuq, fuv, or ful when the precise variety is unknown

Use the row-level language field to select a precise variety. Related languages and varieties within a config should not be assumed interchangeable.

Composition

Config	Rows	Duration	Sources
moore	37,977	98.46 hours	Faso Speech, AfriSpeech, Moore Bible
bambara_jula	93,976	123.33 hours	Faso Speech, AfriSpeech, Koumankan4Dyula, Dyula Speech Bible, merged Bambara-Dioula
fula	49,693	231.22 hours	Faso Speech, Omnilingual ASR, Fula-pular, WaxalNLP, FLEURS

Total: 181,646 rows and 453.01 hours.

Before inclusion, madoss/merged-bambara-dioula-dataset contained 85,040 rows but only 37,848 unique transcripts. Exact comparison found 9,495 transcripts already present in the prior bambara_jula release. After global transcript deduplication, it contributes 28,353 rows.

Minervus00/dyula-speech-bible contains 52,163 source rows and 50,214 unique transcripts. Exact transcript comparison found 1,949 internal duplicate rows and 3 overlaps with the previous bambara_jula release. After global deduplication, it contributes 50,211 rows. The source is gated and does not declare a license, so its rows are recorded with license unknown.

Data splits

Published validation splits are preserved for every source that provides one. Configured upstream dev and test splits are combined into validation, never training. For a source without evaluation data, the builder deterministically holds out --eval-size of that source's training rows using --seed. Final deduplication gives validation rows priority. It uses transcript-only matching for bambara_jula because the merged source lacks per-row language labels, and (language, text) matching for the other configs. The completed build also verifies that no deduplication key remains in both train and validation; the published bambara_jula files have zero transcript overlap between those splits.

Data Fields

Same compact schema as faso-speech, plus per-row provenance:

Column	Description
audio	Embedded Audio(decode=False) value with bytes and path
text	Transcript
language	Precise ISO 639-3 code, a macrolanguage code, or und when the source does not identify Bambara versus Jula
duration	Chunk duration in seconds
id	Source-prefixed native identifier, or a generated 16-character hash
content_type	Source content type, or external when the source doesn't provide one
source	Manifest entry id this row came from (see MANIFEST in the build script)
license	License string recorded for that source in the manifest
attribution	Attribution string recorded for that source in the manifest

Licensing And Attribution

Source licenses differ per entry in the manifest — several are still unknown and must be confirmed from the original source before this dataset is published or used commercially. This is the reason faso-speech-plus is a separate repo from faso-speech: it aggregates third-party data with mixed, not-yet-fully-verified licensing, unlike the curated mooreburkina.com-sourced faso-speech corpus.

Consult the row-level license, source, and attribution fields before use. Some included sources do not declare a license and are recorded as unknown; users are responsible for verifying that their intended use is permitted.