Introduction

Tɛkyerɛma Dagbani ASR just shipped as an open Hugging Face checkpoint: PrinceAlhassanNasamu/tekyerema-asr-dag (created 2026-11-04 ~22:31 UTC). It is the Dagbani (dag) hearing model behind the Tɛkyerɛma / Nexus language service — a CTC fine-tune of Khaya AI’s multi-Gur w2v-BERT 2.0 on University of Ghana WAXAL speech, licensed CC BY 4.0.

If you build low-resource West African speech tools, this is a rare ungated Dagbani card with a real held-out score and a one-liner transformers path.

What Shipped

  • Vendor / product: Tɛkyerɛma Dagbani ASR v1 — Prince Nasamu Alhassan (Tɛkyerɛma / Nexus)
  • Modality: open-weight automatic speech recognition (Dagbani)
  • Open vs closed: open weights, CC BY 4.0 (base Khaya model Apache-2.0; WAXAL data CC-BY-4.0)
  • Runnable today: transformers ASR pipeline on the HF repo (CPU or GPU)
  • Concrete fact: fine-tuned 2,751 steps on 9,273 WAXAL Dagbani clips (45.3 h); held-out WAXAL test WER 0.872 → 0.358 and CER 0.335 → 0.122 (150 clips each); root model.safetensors is about 2.42 GB

A second check on GhanaNLP bible_Dagbani (ghana-speech-eval) moves more gently: WER 0.684 → 0.601, CER 0.253 → 0.210 — still useful as an out-of-domain sanity bar.

Run It

pip install transformers torch torchaudio soundfile

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="PrinceAlhassanNasamu/tekyerema-asr-dag",
)
print(asr("clip16k.wav")["text"])

Feed 16 kHz mono wav when you can. The card’s base is KhayaAI/w2v-bert-gjn_maw_gur_dag_dga_kus_lxn_wlx_xon_xsm_en, so you inherit a Gur-family encoder already exposed to Dagbani and related languages before the CTC head was specialized.

Notes for makers

  • Credit Khaya AI, the WAXAL / WaxalNLP project and University of Ghana contributors, plus Tɛkyerɛma (Prince Nasamu Alhassan) — CC BY wants attribution
  • Expect a large first download (~2.4 GB safetensors); cache it if you iterate locally
  • Domain shift still bites (bible eval is softer than WAXAL test) — treat scores as guidance, not a production SLA
  • Soft-hold neighbors like GhanaOpenAI ZipVoice / ghana-fast-voice remain 401 or NC / unpackaged; this Dagbani ASR is the ungated Ghana-adjacent scoop this cycle
  • No Gen CTA: low-resource ASR does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

For Dagbani makers who need an open, attributable ASR checkpoint they can run from Hugging Face today, Tekyerema’s WAXAL fine-tune is the zero-day card — pull it, credit the chain, and measure on your own domain before you ship.

—Anabel ♡