Whisper-Small-Shona 🇿🇼

The first open Shona (chiShona) speech recognition model trained on real Shona speech data.

Part of Project Nyaradzai — open language infrastructure for chiShona and isiNdebele.

Model description

This model is a fine-tuned version of openai/whisper-small for automatic speech recognition (ASR) in chiShona, the primary language of Zimbabwe spoken by over 10 million people.

It was trained on the Google WAXAL dataset (google/WaxalNLP, config sna_asr) — a large-scale, CC-BY-4.0 licensed multilingual African speech corpus. This is currently the largest openly available Shona speech dataset.

Property	Detail
Base model	openai/whisper-small
Language	chiShona (sn)
Task	Automatic Speech Recognition
Training data	google/WaxalNLP (sna_asr split)
Training examples	~13,000 labelled utterances
Training steps	4,000
Word Error Rate	36.42% (WAXAL test split)
License	Apache-2.0

Why this model matters

ChiShona has been effectively absent from speech technology. Speakers cannot:

• Caption or transcribe Shona TV and video
• Use voice-to-text in Shona on phones or computers
• Access accessibility tools (subtitles, screen readers) in their language

This model is a first step toward fixing that. A 36.42% WER on a first fine-tune with no Zimbabwean broadcast-domain data is a solid foundation — improving with more data is the next step.

How to use

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="Starsm91/whisper-small-shona"
)

# Transcribe an audio file
result = asr(
    "your_shona_audio.mp3",
    generate_kwargs={"language": "shona", "task": "transcribe"},
    return_timestamps=True
)
print(result["text"])

For audio clips shorter than 30 seconds you can omit return_timestamps:

result = asr(
    "short_clip.mp3",
    generate_kwargs={"language": "shona", "task": "transcribe"}
)
print(result["text"])

Test on WAXAL data directly

from transformers import pipeline
from datasets import load_dataset

asr = pipeline("automatic-speech-recognition",
               model="Starsm91/whisper-small-shona")

ds = load_dataset("google/WaxalNLP", "sna_asr",
                  split="test", streaming=True)

for i, item in enumerate(ds):
    if i >= 3:
        break
    result = asr(item["audio"]["array"],
                 generate_kwargs={"language": "shona", "task": "transcribe"},
                 return_timestamps=True)
    print(f"Model:     {result['text'].strip()}")
    print(f"Reference: {item['transcription'].strip()}\n")

Training details

Parameter	Value
Learning rate	1e-5
Batch size	4 (with gradient accumulation of 4)
Warmup steps	200
Max steps	4,000
FP16	Yes
Framework	HuggingFace Transformers
Hardware	Kaggle T4 GPU
Training time	~5 hours

Data preprocessing

• Loaded train split only (~14,109 examples)
• Filtered examples with missing transcripts or audio shorter than 0.1 seconds
• 95/5 train/eval split (seed=42)
• Audio resampled to 16kHz

Evaluation

Word Error Rate (WER): 36.42% on the WAXAL test split.

This means the model correctly transcribes roughly 2 in every 3 Shona words. For context:

• This is a first fine-tune with no Zimbabwean broadcast-domain data
• The WAXAL data covers diverse Shona speakers from multiple regions
• Expected improvement with: more training steps, Zimbabwean broadcast audio, Common Voice data

Limitations

• Trained exclusively on WAXAL speech — may struggle with strong regional Zimbabwean accents (Harare urban, rural dialects)
• No broadcast/news domain data — radio and TV speech may score higher WER
• 36.42% WER means roughly 1 in 3 words will need correction for formal use
• Best suited for: voice memos, conversational transcription, accessibility tools, research

Roadmap

The next planned improvements (via Project Nyaradzai):

1. v0.6 — Fine-tune on Zimbabwean broadcast audio (ZBC radio archives, community radio)
2. Common Voice campaign — crowdsource Shona recordings across Zimbabwe for a locally-representative dataset
3. whisper-medium-shona — larger model for higher accuracy once more data is available
4. Live captioning pipeline — deploy for Zimbabwean broadcasters

Citation

If you use this model in your research or work, please cite:

@misc{whisper-small-shona-2026,
  title={Whisper-Small-Shona: First Open Shona ASR Model},
  author={Mateta, Stanley},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Starsm91/whisper-small-shona},
  note={Part of Project Nyaradzai: https://github.com/stanleymateta-tech/Project-Nyaradzai}
}

Data attribution

Training data: Google WAXAL (West African and eXtra-continental African Languages) dataset. License: CC-BY-4.0. Citation: WAXAL: A Multilingual Speech Dataset for African Languages, Google Research, 2026.

Project Nyaradzai

This model is part of a larger open-source project to bring chiShona and isiNdebele into the digital era:

• 🔤 Spellchecker: 214,000+ word forms for Word, LibreOffice, Firefox
• 🗣 ASR model: This model
• ⌨️ Android keyboard: In development
• 📚 Validated lexicon: Community-driven, in progress

GitHub: github.com/stanleymateta-tech/Project-Nyaradzai

Contributions welcome — especially from Shona and Ndebele native speakers. No coding required to contribute: report a word.

---

Mutauro wedu, panyika yose — Our language, everywhere in the world.