README.md

---
language:
  - men
license: apache-2.0
library_name: transformers
pipeline_tag: automatic-speech-recognition
tags:
  - speech
  - asr
  - african-languages
  - w2v-bert
  - khaya
  - men
  - arxiv:2607.21540
metrics:
  - wer
model-index:
  - name: w2v-bert-men
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        metrics:
          - type: wer
            value: 12.9
            name: WER
---

![dondo_v1_graphic](https://cdn-uploads.huggingface.co/production/uploads/5f49548279c1ba4c353d122c/NMp40RI78sjS1fWQUzlUc.png)
# w2v-bert-men — Mende ASR

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

- **Language:** Mende (`men`)
- **Task:** Automatic Speech Recognition (CTC)
- **Backbone:** w2v-BERT 2.0
- **License:** Apache-2.0 (attribution only; commercial use permitted)
- **Paper:** [arXiv:2607.21540](https://arxiv.org/abs/2607.21540)
- **Demo:** [Try it in your browser](https://huggingface.co/spaces/Ghana-NLP/Sierra-Leone-ASR)
- **Part of:** the DONDO family — <https://huggingface.co/KhayaAI>

## Intended use

A **base model** for Mende 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) | **12.9%** |

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-men"
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.