License production-grade Shona and Ndebele speech corpora. Benchmark your models on the only public leaderboards for Zimbabwean languages. Or get paid to build the dataset yourself.
Every release is a complete handover package: audio, word-aligned JSONL transcripts, machine-readable schema, validation reports, bias register and consent documentation. The loading script is included — running it is our release criterion, not your integration problem.
# pandas + librosa. No manual edits. That's the point. import pandas as pd, librosa utts = pd.read_json("processed/utterances.jsonl", lines=True) wav, sr = librosa.load(utts.audio_file[0], sr=None) assert (utts.consent_ref.notna()).all() # always true assert (utts.qa_status == "pass").all() # always true # 47 sessions · 96 channels · word-level alignment # dialects: zezuru · karanga · manyika · korekore
Every corpus releases a free CC BY 4.0 subset to the commons — with competitions, held-out test sets and public leaderboards. Students and developers stop practising on foreign data about foreign problems. Institutions bring real problems and sponsor the next season.
Submitted solutions, evaluation results and benchmarks flow back into the commons — the competition itself becomes a dataset.
Enter the benchmarkRecord conversations, read scripts, review translations — from an entry-level Android phone, on constrained networks, with resumable uploads. Published per-task rates, foreign-currency payouts over mobile money, and a ledger you can audit yourself. Contributor payments are the first call on revenue, at a fixed published share.
A recording without a valid consent link isn't repaired — it's rejected, automatically, by the pipeline. Zimbabwe's Data Protection Act [Chapter 12:07] is the floor, not the ceiling.