eko-stt-v1

State-of-the-art speech recognition for Nigerian English, Nigerian Pidgin, Yorùbá and Igbo, by Ekogenie. A 1.55B-parameter encoder-decoder recogniser, and the model behind Ekogenie's live voice agents.

Try it in your browser: eko-stt-v1 live demo. Speak or upload a clip and pick a language.

Headline

Compared with SBPN (paper, weights), the previous published state of the art for these four languages, on the same 16 test sets with the same scorer:

• Mean word error rate 13.9% vs 18.4% (24% relative reduction).
• Better on 14 of 16 test sets: every Nigerian English, Pidgin and Igbo set, and 2 of 4 Yorùbá sets.
• Pidgin error roughly halved (for example PidginMed 7.3 vs 13.5, WaZoBia 9.5 vs 17.8); Nigerian English AfriSpeech 9.3 vs 16.4.
• SBPN-large is still ahead on Common Voice Yorùbá and NaijaVoices Yorùbá. Both are corpora SBPN trained on.

Language	eko-stt-v1	SBPN-large
Nigerian English	7.5	14.0
Nigerian Pidgin	9.3	16.9
Igbo	18.3	21.0
Yorùbá	20.3	20.8

Results

Word error rate in %, lower is better. Both models are scored on identical clips by the same scorer.

• Normalised: the reference and the hypothesis both go through the same normaliser: lowercase, no punctuation, numbers in one form, fillers dropped, hyphens split, documented spelling equivalences (including English and Naija spellings of the same Pidgin word). For Yorùbá and Igbo, tone marks are ignored and dot-below is kept, because published references mark tone inconsistently.
• Raw: lowercase and punctuation only.
• With tone: normalised, tone marks counted, against test references whose tone marks (Yorùbá) and dot-below (Igbo) were unified by a text-only pass (keep existing marks, add missing ones, letters never changed). Our model's output was not shown to that pass, so it does not favour this model.

Test set	Language	Clips	Ours, normalised	SBPN-large, normalised	Ours, raw	SBPN-large, raw	Ours, with tone	SBPN-large, with tone
AfriSpeech-200 Nigerian test, reviewed	Nigerian English	1517	9.3	16.4	12.0	20.3	-	-
AfriSpeech-200 Nigerian test	Nigerian English	1343	10.6	17.8	13.3	21.8	-	-
Common Voice 27 Igbo test	Igbo	500	19.4	21.6	27.9	29.5	21.1	23.5
Common Voice 27 Pidgin test	Nigerian Pidgin	326	16.4	26.3	22.7	33.9	-	-
Common Voice 27 Yorùbá test	Yorùbá	500	18.5	16.1	25.8	22.7	20.7	18.4
DSN African Voices Igbo devtest	Igbo	500	14.7	18.3	24.9	26.6	17.4	20.9
DSN African Voices Yorùbá devtest	Yorùbá	500	20.2	25.9	31.1	41.8	28.3	30.8
FLEURS Igbo test	Igbo	869	22.1	26.8	32.0	34.6	22.2	26.9
FLEURS Yorùbá test	Yorùbá	817	23.4	24.7	44.6	43.7	34.8	32.8
NaijaVoices Igbo official test	Igbo	1006	16.9	17.5	26.8	25.2	17.5	18.2
NaijaVoices Yorùbá official test	Yorùbá	1142	19.1	16.4	34.6	32.2	24.1	21.0
Native Nigerian English (100 clips, held out)	Nigerian English	100	2.5	7.6	3.9	8.3	-	-
PidginMed test	Nigerian Pidgin	976	7.3	13.5	8.8	20.3	-	-
PidginMed test, reviewed	Nigerian Pidgin	976	6.7	13.0	8.2	19.8	-	-
WaZoBia Pidgin devtest, clean	Nigerian Pidgin	500	9.5	17.8	14.7	26.4	-	-
WaZoBia Pidgin devtest (300)	Nigerian Pidgin	300	6.4	13.9	10.9	21.5	-	-

Mean over the 16 sets (normalised): 13.9 vs 18.4.

Usage

from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="munasco/eko-stt-v1", torch_dtype="float16", device="cuda")
asr("clip.wav", generate_kwargs={"language": "yo", "task": "transcribe"})

Language tokens: Nigerian English and Pidgin use en, Yorùbá uses yo, and Igbo uses the repurposed ln token (the base vocabulary has no Igbo token). Leaving the language unset (auto-detect) performs about the same as forcing it. Use greedy decoding for live use (beam 5 gains about 0.5 point at 6x the latency), and a repetition guard on long audio.

Training

• Full fine-tune of an open-weights (Apache-2.0) multilingual base, 13,000 updates, checkpoint chosen on a held-out dev mix.
• About 1,070 hours of transcribed Nigerian speech across the four languages: read speech, broadcast media and conversational Pidgin. No synthetic audio.

Limitations

• Yorùbá tone marking is the main remaining error. With unified references and tone counted, Yorùbá error is about 8 points higher than without tone.
• Names and rare words are the next most common error.
• Hausa is not covered yet.

Contact

Munachi Ernest-Eze, Ekogenie, munachi@ekogenie.com