DONDO (Democratizing Oral Neural Dialect Ontology) open speech-recognition base model for Nzema (Ghana). Fine-tuned from the w2v-BERT 2.0 self-supervised speech encoder.

Language: Nzema (nzi)
Task: Automatic Speech Recognition (CTC)
Backbone: w2v-BERT 2.0
License: Apache-2.0 (attribution only; commercial use permitted)
Paper: arXiv:2607.21540
Demo: Try it in your browser
Part of: the DONDO family — https://huggingface.co/KhayaAI

Intended use

A base model for Nzema ASR. Use it directly to transcribe read or relatively clean speech, or as an initialisation to fine-tune on your own in-domain data (conversational, broadcast, clinical, etc.) with comparatively little labelled audio.

DONDO is also intended as an open test bed: a shared, reproducible model on which new low-resource speech techniques can be demonstrated for the benefit of all. Aligned research groups are welcome to build on it and collaborate.

Training data

Primarily read speech derived from religious texts paired with verified transcripts in the standard orthography. Such data is license-clear, exists for many otherwise under-resourced languages, and is orthographically consistent. Its main limitation is domain narrowness, which is why this is released as a base model.

Evaluation

Metric	Value
WER (in-domain test)	12.6%

WER is computed on in-domain test material and should not be read as a guarantee for other domains.

How to use

python

import torch, torchaudio
from transformers import AutoProcessor, AutoModelForCTC

model_id = "KhayaAI/w2v-bert-nzi"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id)

speech, sr = torchaudio.load("audio.wav")
if sr != 16000:
    speech = torchaudio.functional.resample(speech, sr, 16000)
inputs = processor(speech.squeeze().numpy(), sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
pred_ids = torch.argmax(logits, dim=-1)
print(processor.batch_decode(pred_ids)[0])

Limitations

Trained largely on read religious speech, so the model may underperform on spontaneous, code-switched or noisy audio until fine-tuned. Orthographic conventions vary across communities; evaluation for the smallest languages rests on limited test sets.

License

Released under the Apache-2.0 license. You are free to use, modify, redistribute and build upon this model, including for commercial purposes. The only substantive requirement is attribution.

Professional services & hosted APIs

This model is free to use under Apache-2.0. If your team would like help deploying it offline on your own infrastructure and data, Khaya AI offers professional services (integration, fine-tuning and on-premises deployment). For ready-to-use and more advanced ASR, including APIs for interested parties, see Khaya Studio.

Khaya AI — https://khaya.ai
Khaya Studio — https://studio.khaya.ai
Khaya Studio ASR — https://studio.khaya.ai/asr

Citation

bibtex

@article{azunre2026dondo,
  title         = {DONDO: Open w2v-BERT Speech Recognition Base Models for African Languages},
  author        = {Azunre, Paul and Ibrahim, Naafi and Budu, Joel and Adu-Gyamfi, Lawrence},
  year          = {2026},
  eprint        = {2607.21540},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  note          = {Democratizing Oral Neural Dialect Ontology. Funded by the Huniki Federation.}
}

Acknowledgements

Funded by the Huniki Federation. We thank Ghana-NLP and Algorine Research for their support with benchmarking, testing and data, and Hugging Face for compute credits. We also thank the language communities and data contributors whose recordings and transcriptions made this work possible.