A speech corpus for Shona, built with the people who speak it
Inzwi (“voice” in Shona) collects consented, peer-reviewed, dialect-tagged recordings of Shona speech, and uses them to measure how much better speech recognition gets when it is trained on them. Ndebele follows.
The gap this research addresses
Speech recognition still works poorly for Shona. The transcribed Shona speech that exists is mostly one style of speaking, carries no dialect labels, and was not gathered under Zimbabwean governance.
That matters because Shona is not one voice. Zezuru, Karanga, Manyika and Korekore speakers pronounce and choose words differently, and a model that has only heard one of them will fail the others. This research builds a corpus where every clip says which dialect it is, and where that label has been checked by other speakers of the same dialect.
How the data is collected
Contributors take part on WhatsApp or on this site. Every step below is part of the documented method.
Prompted reading
Contributors read a Shona sentence aloud. The sentence is the transcript, so every clip arrives already labelled. Prompts are written to cover numbers, names, code-switching and sounds that are easy to confuse.
Free talk
Contributors answer an everyday question in their own words. This captures the natural rhythm of speech that reading cannot. These answers are transcribed afterwards.
Peer review
Other contributors listen to each clip. A clip is accepted or rejected only when two reviewers agree, and every rejection records a reason from a fixed list, so the causes can be counted.
Dialect verification
Contributors declare their dialect. Reviewers who share it are asked whether the speaker sounds like it, so the dialect label on each clip is confirmed by other speakers, not only self-reported.
What the research produces
The corpus
Validated, dialect-tagged Shona speech with transcripts, contributor consent and the review record behind every clip.
The method
A documented, repeatable way to crowdsource speech over WhatsApp and the web in a low-bandwidth setting — review, dialect checks and withdrawal included.
A measured result
A Whisper speech-recognition model fine-tuned on the corpus, with its word error rate compared against the untuned model on the same test set.
Your recordings
- You agree to take part before your first recording, and you can stop at any time.
- Recordings are never public links. They are served only on signed, expiring addresses to people reviewing them.
- You can withdraw your consent from Settings → Privacy, or by sending "withdraw" on WhatsApp. Your recordings and transcripts are then deleted and your contact details cleared.
- Your name and phone number are never part of the dataset. Clips carry an anonymous speaker code, your dialect, age range and gender, and the text that was read.
The full terms are in the privacy policy.
Questions about the research
About taking part, your data, or working with the corpus.
hello@inzwi.app