Datasetpublic

ghanaopenai/ghana-speech-eval

ghana-speech-eval ⚠️ Evaluation only. This is a held-out benchmark for measuring ASR quality on Ghanaian languages. Please do not train or fine-tune on it — doing so contaminates the benchmark and makes reported scores meaningless. Every subset ships a single eval split. Multi-source ASR evaluation benchmark for Ghanaian languages. All subsets are 16 kHz mono with standardised schema (audio, text, language, country, length, iso, subset). Source groups… See the full description on the dataset page: https://huggingface.co/datasets/ghanaopenai/ghana-speech-eval.

sourceHugging Faceupdated 29d agoView on Hugging Face

0likes581downloads

Dataset Card

ghana-speech-eval

⚠️ Evaluation only. This is a held-out benchmark for measuring ASR quality on Ghanaian languages. Please do not train or fine-tune on it — doing so contaminates the benchmark and makes reported scores meaningless. Every subset ships a single eval split.

Multi-source ASR evaluation benchmark for Ghanaian languages. All subsets are 16 kHz mono with standardised schema (audio, text, language, country, length, iso, subset).

Source groups

Group	Description	Clip length	Per-language cap
`bible_*`	Bible-audio aligned (ghananlpcommunity/ghana-speech)	3–15 s	1 000
`finance_*`	Finance-domain multispeaker speech (ghanaopendata)	0–60 s	1 000
`jw_*`	JW.org speech (AfriSpeech v1)	3–15 s	1 000
`lds_*`	LDS general-conference speech (multispeaker)	3–15 s	1 000
`unicef_*`	UNICEF health recordings	3–60 s	all available (~195–200)
`waxal_*`	WaxalNLP ASR (google/WaxalNLP)	3–30 s	1 000

Twi is split by dialect

Akuapem and Asante Twi are separate languages here, twi_akuapem and twi_asante. They were previously merged into a single twi (500/500 in bible_twi_twi and finance_twi), which made every Twi score an average over two orthographies: the Khaya API's Akuapem endpoint scored 8.9 % WER on the Akuapem-heavy bible split and 88.7 % on jw, and its general-Twi endpoint the reverse. The merged configs (bible_twi_twi, finance_twi, unicef_twi, jw_twi_twi) have been removed. jw_* remains for the other nine languages — its Twi audio mixes both dialects with no way to separate them.

Neither dialect has its own ISO 639-3 code (both are twi), so twi_akuapem and twi_asante are this dataset's keys, not ISO.

Two labels are our classification rather than the source's, and are stated here so they are not mistaken for ground truth:

—lds_Asante_Twi — the source is labelled only "Twi (Akan)"; its orthography (mfeɛ, afoforɔ) matches the verified Asante bible text.
—waxal_Asante_Twi — WaxalNLP labels this config aka / "Akan" without a dialect; it is filed under Asante Twi.

Configs (65 total)

`bible_*` — 42 configs

Config	ISO	Language	Clips
`bible_Gikyode_acd`	acd	Gikyode	1000
`bible_Dangme_ada`	ada	Dangme	1000
`bible_Siwu_akp`	akp	Siwu	1000
`bible_Anyin_any`	any	Anyin	1000
`bible_Avatime_avn`	avn	Avatime	1000
`bible_Bissa_bib`	bib	Bissa	1000
`bible_Bimoba_bim`	bim	Bimoba	1000
`bible_Birifor_Southern_biv`	biv	Birifor_Southern	1000
`bible_Tuwuli_bov`	bov	Tuwuli	1000
`bible_Bassar_Ntcham_bud`	bud	Bassar_Ntcham	1000
`bible_Buli_bwu`	bwu	Buli	1000
`bible_Dagbani_dag`	dag	Dagbani	1000
`bible_Dagaare_dga`	dga	Dagaare	1000
`bible_Ewe_ewe`	ewe	Ewe	1000
`bible_Fante_fat`	fat	Fante	1000
`bible_Fulfulde_Maasina_ffm`	ffm	Fulfulde_Maasina	1000
`bible_Gonja_gjn`	gjn	Gonja	1000
`bible_Ninkare_gur`	gur	Ninkare	1000
`bible_Hausa_hau`	hau	Hausa	1000
`bible_Kabiye_kbp`	kbp	Kabiye	1000
`bible_Tem_kdh`	kdh	Tem	1000
`bible_Konni_kma`	kma	Konni	1000
`bible_Kusaal_kus`	kus	Kusaal	1000
`bible_Lelemi_lef`	lef	Lelemi	1000
`bible_Sekpele_lip`	lip	Sekpele	1000
`bible_Mampruli_maw`	maw	Mampruli	1000
`bible_Deg_mzw`	mzw	Deg	1000
`bible_Nawuri_naw`	naw	Nawuri	1000
`bible_Chumburung_ncu`	ncu	Chumburung	1000
`bible_Nkonya_nko`	nko	Nkonya	1000
`bible_Ntrubo_ntr`	ntr	Ntrubo	1000
`bible_Nzema_nzi`	nzi	Nzema	1000
`bible_Sehwi_sfw`	sfw	Sehwi	1000
`bible_Paasaal_sig`	sig	Paasaal	1000
`bible_Sisaala_Tumulung_sil`	sil	Sisaala_Tumulung	1000
`bible_Selee_snw`	snw	Selee	1000
`bible_Tampulma_tpm`	tpm	Tampulma	1000
`bible_Akuapem_Twi`	twi_akuapem	Akuapem Twi	1000
`bible_Asante_Twi`	twi_asante	Asante Twi	1000
`bible_Vagla_vag`	vag	Vagla	1000
`bible_Konkomba_xon`	xon	Konkomba	1000
`bible_Kasem_xsm`	xsm	Kasem	1000

`finance_*` — 4 configs

Config	ISO	Language	Clips
`finance_fante`	fat	Fante	1000
`finance_ga`	gaa	Ga	1000
`finance_Akuapem_Twi`	twi_akuapem	Akuapem Twi	1000
`finance_Asante_Twi`	twi_asante	Asante Twi	1000

`jw_*` — 9 configs

Config	ISO	Language	Clips
`jw_dangme_ada`	ada	Dangme	261
`jw_ahanta_aha`	aha	Ahanta	1000
`jw_dagaare_dga`	dga	Southern Dagaare	1000
`jw_ewe_ewe`	ewe	Éwé	1000
`jw_fante_fat`	fat	Fante	1000
`jw_ga_gaa`	gaa	Ga	1000
`jw_frafra_gur`	gur	Farefare	1000
`jw_nzema_nzi`	nzi	Nzema	1000
`jw_sehwi_sfw`	sfw	Esahie	1000

`lds_*` — 2 configs

Config	ISO	Language	Clips
`lds_Fante_fat`	fat	Fante	1000
`lds_Asante_Twi`	twi_asante	Asante Twi	1000

`unicef_*` — 3 configs

Config	ISO	Language	Clips
`unicef_dagbani`	dag	Dagbani	195
`unicef_ewe`	ewe	Éwé	199
`unicef_Asante_Twi`	twi_asante	Asante Twi	196

`waxal_*` — 5 configs

Config	ISO	Language	Clips
`waxal_Dagbani_dag`	dag	Dagbani	1000
`waxal_Dagaare_dga`	dga	Dagaare	1000
`waxal_Ewe_ewe`	ewe	Ewe	1000
`waxal_Ikposo_kpo`	kpo	Ikposo	1000
`waxal_Asante_Twi`	twi_asante	Asante Twi	1000

Loading

python

from datasets import load_dataset

# every subset has a single `eval` split
ds = load_dataset("ghananlpcommunity/ghana-speech-eval", "bible_Asante_Twi", split="eval")

from datasets import get_dataset_config_names
print(get_dataset_config_names("ghananlpcommunity/ghana-speech-eval"))

Schema

Column	Type	Description
`audio`	Audio	16 kHz mono PCM
`text`	string	reference transcription (verbatim)
`language`	string	human-readable language label
`country`	string	ISO 3166-1 alpha-2, or ""
`length`	float64	clip duration in seconds
`iso`	string	language key (ISO 639-3, except the two Twi dialect keys)
`subset`	string	`<group>_<label>`, matching the config name

Built with speecheval-builder. Benchmarked by nsanku-ASR.

License

CC-BY-4.0. Respect the terms of each source dataset.

ghana-speech-eval

Source groups

Twi is split by dialect

Configs (65 total)

bible_* — 42 configs

finance_* — 4 configs

jw_* — 9 configs

lds_* — 2 configs

unicef_* — 3 configs

waxal_* — 5 configs