sable/shona-duplex-conversation — Sable Foundry — Light
datasets / sable / shona-duplex-conversation

Shona Full-Duplex Conversational Corpus

v1.0.0 ● released
Two-speaker unscripted conversations, one channel per speaker — overlap, interruptions and backchannels preserved. The first commercial speech corpus for Shona.
size 10 hrs · 47 sessions format WAV PCM 48 kHz + JSONL licence commercial · 20% open subset updated 2026-07-02 steward Tendai Gatahwa
{{ t.label }}
Sample sessions 30-second preview clips · pseudonymised
▶ {{ s.id }}
{{ s.meta }} {{ s.dur }}
{{ k.label }} {{ k.value }}
What's in the box 96 speaker-channel WAV files (one per speaker, overlap preserved) · word-aligned UTF-8 JSONL transcripts with dialect, event and code-switch tags · metadata cards · schema + data dictionary · validation reports and error log · bias register · consent documentation · load_corpus.py
METADATA CARD
{{ m.k }} {{ m.v }}
FREE OPEN SUBSET 2 hours under CC BY 4.0 — no licence needed for education and research. Get sable/shona-duplex-open →
FIELDTYPEREQDESCRIPTION / ALLOWED VALUES
{{ f.name }} {{ f.type }} {{ f.req }} {{ f.desc }}
Quality report — v1.0.0
{{ q.label }} {{ q.value }}
How it was checked Automated gates on every submission (integrity, clipping, SNR, silence, duplicates, schema), then human QA for speaker separation, topic compliance and consent linkage. Transcripts: machine first pass, corrected by a trained transcriber, with an independent second review of a 15% sample per transcriber.
Release criterion A sample of the release must load into Python (librosa / pandas) without manual editing. Any failed check blocks release — pass/warning/fail conditions ship in validation_rules.json.
Disclosed coverage — v1.0.0 gaps are disclosed, never concealed
DIMENSIONTARGETACTUALNOTE
{{ b.dim }} {{ b.target }} {{ b.actual }} {{ b.note }}
CONSENT STATUS Every utterance is consent-linked All 60+ speakers gave informed, per-purpose, revocable consent in their own language, evidenced by per-submission consent records. Linkage completeness is 100% by construction — the pipeline rejects anything unlinked. The consent register is audited as part of this release's validation report.
LICENCE PROHIBITIONS What you may not do
✕ Speaker identification or surveillance of any kind ✕ Credit / insurance decisioning against individuals ✕ Law-enforcement profiling ✕ Impersonation of identifiable voices ✕ Re-sale or re-identification attempts
contractual · audit rights · termination clauses