languages
Transcribe Shona audio to text
Transcribe Shona (chiShona) recordings on your device. Start with a short sample and use the paper’s results below to set realistic expectations.
Last checked:
chiShona / FLEURS WER
126%
base model · Lowest published error rate among the models offered here.
Original Whisper paper, Table 13. Lower is better. These are not browser-model test results. Source and context ↓
Tiny transcription · 40.8 MB CPU / 119.7 MB GPU weights. Runtime files download too.
Choose a file or record your microphone.
Download and processing settings
Model weights: 40.8 MB for WebAssembly; 119.7 MB for GPU. Configuration, tokenizer and runtime files add a smaller download. Source: Hugging Face file sizes, checked 2026-10-30.
Model file sizes and source · Last checked October 30, 2026
Tiny uses the least memory. Base and Small may improve recognition but need more memory and time. Try a short sample first. No accuracy or speed is guaranteed.
Audio is processed locally. Model files download from Hugging Face when you start, then stay in your browser cache when storage permits. How privacy works
Your transcript
Click a word or timestamp to listen. Search updates as you type. Edit text and add speaker names yourself. Speaker recognition is not automatic.
Shape your subtitles
SRT and VTT exports use this layout. These controls limit characters and lines, not reading speed. Check subtitle reading speed before publishing.
Saved on this device
Recordings and edits are stored in this browser when local storage is available. Export an important transcript before clearing browser data. The delete control clears app storage and its model cache. Use browser cache controls to remove any browser-managed download copies too.
Shona model results
These are original Whisper paper FLEURS results, not tests of MurmurScribe or its quantized browser models. Lower error rates are better. The paper counts substitutions, deletions and insertions against the reference words after its normalization procedure.
| Model | FLEURS WER (%) |
|---|---|
| tiny | 128.1 |
| base | 126 |
| small | 140 |
Source: Whisper paper, Table 13. No corresponding Common Voice result is included in this extraction. An absent column is not an error rate of zero.
Choose a starting model
Among the models offered here, base has the lowest published FLEURS WER for Shona, at 126%. Small has a higher error rate than Tiny on this test. Increasing model size does not guarantee an improvement for this language. This comparison is a starting point, not a prediction of your finished transcript.
The published errors are very high. Treat automatic output as an uncertain draft, and consider manual transcription or another model if a representative sample is unusable. Do not infer fluency or language coverage from the presence of a language option.
Training exposure and review
A separate Shona speech-recognition training-hour label is not present in the extracted Figure 11 panel. We leave this unknown rather than treating an absent label as zero. Training volume alone does not describe accents, domains or recording quality.
Set the audio language to Shona when you know it. Keep the task on transcription to preserve the spoken language; translation produces English instead. Review names, borrowed words and changes of speaker against the recording. Export the corrected text only after checking passages you plan to quote. Keep uncertain phrases marked until a fluent listener can resolve them.
Frequently asked questions
Can I transcribe Shona without uploading audio?
Yes. The browser downloads the model and runs it locally. Audio and transcript contents are not uploaded.
Do these scores guarantee my accuracy?
No. These are original paper benchmark results. Quantization, recording conditions and the chosen language affect the draft.
Can I translate the recording into English?
Choose Translate to English in the app. Have a fluent person check meaning and names before publishing.
Sources
- Whisper paper, appendix tables and normalization method · OpenAI / Radford et al. · Checked 2026-10-30
- Whisper model details and MIT license · OpenAI · Checked 2026-10-30
- Browser model execution · Hugging Face · Checked 2026-10-30