nigeLbasa/inzwi-shona-sample

Inzwi — Shona Speech Corpus (Cleaned Sample)

A small, fully-documented sample of the Inzwi speech corpus: consented, peer-validated Shona audio paired with ground-truth transcripts, prepared to be AI-ready for automatic speech recognition (ASR). Built for the POTRAZ AI for Impact (AI4I) Challenge — Data Track, to demonstrate the Inzwi data pipeline end to end.

This is a representative sample (the validated slice of an early, un-incentivised run), not the full corpus. It exists to show how Inzwi produces trustworthy, machine-readable local-language data — the quality of the pipeline, not the size of the pile.

Live platform: https://inzwi.app · Contribute on WhatsApp: https://wa.me/+263719987483

What's in this sample

Clips	42
Speakers	6 (pseudonymised SPK01–SPK06)
Total audio	~1.85 min (110.8 s)
Content	14 single words + 28 sentences
Language	Shona (sn)
Dialects	Zezuru (40), Korekore (2)
Audio format	WAV, 16 kHz, mono, 16-bit PCM (ASR-ready)
Transcripts	transcripts.jsonl (one JSON object per clip)
Speaker metadata	pseudonymous speaker_id + optional self-reported gender (19/42; 10 M / 9 F) for coverage balancing. age_range optional — not provided in this slice
Per-clip quality signals	avg_volume_db, peak_volume_db, silence_percentage, has_clipping, noise_level, device_type, file_size_bytes, validation_score, status
Licence	Transcripts + metadata CC BY 4.0; audio is a consented public sample

v1.1.0 (metadata enrichment): same 42 clips as v1.0.0, now with full machine-ready per-clip metadata and optional aggregate gender. See data_dictionary.csv for every field.

The Inzwi data pipeline (raw → AI-ready)

CONTRIBUTOR                 INGEST                 PRE-PROCESS (automatic)          VALIDATE            EXPORT
  Web PWA        ─┐                               ├ normalise → 16 kHz mono WAV                     ┌ transcripts.jsonl
  WhatsApp bot   ─┼─►  prompted reading  ─►  raw audio  ─► silence-trim              ─► peer review ─┼ WAV audio
  (Meta Cloud)   ─┘    (text = ground-      (.ogg/.webm)  ─► SNR / volume scoring       (accept/       ├ metadata + manifest
                        truth transcript)                 ─► clipping detection          reject)       └ data card

1. Sources. First-party voluntary contributors record prompted readings — they read a curated Shona sentence, so the prompt text is the ground-truth transcript (labels correct by construction). Two channels: web PWA and a WhatsApp bot (Meta Cloud API).
2. Ingest. Raw audio lands as .ogg/.webm with contributor + prompt linkage.
3. Pre-processing (automatic, in the background). Format normalisation to 16 kHz mono WAV, silence trimming, signal-to-noise and volume scoring, and clipping detection. Recordings below quality thresholds are flagged.
4. Validation. Human peer review accepts or rejects each clip; a pending → validated / rejected lifecycle with a recorded validation state. Only validated clips appear here.
5. Export. Delivered as WAV + transcripts.jsonl + machine-readable metadata, manifest (with checksums), and this data card — loadable into an ASR pipeline with no manual cleanup.

Files

data/train/metadata.csv  # drives the HF dataset viewer (playable audio table)
data/train/*.wav         # 42 clips, 16 kHz mono WAV
transcripts.jsonl        # audio↔text pairs + metadata, one JSON object per line
metadata.json            # dataset-level metadata card (machine-readable)
data_dictionary.csv      # field definitions for transcripts.jsonl
manifest.csv             # every file: path, bytes, sha256, duration, sensitivity

Load it (Python)

import json, soundfile as sf
rows = [json.loads(l) for l in open("transcripts.jsonl", encoding="utf-8")]
audio, sr = sf.read(rows[0]["audio"])   # 16 kHz mono, ready for Whisper/wav2vec2
print(rows[0]["text"], sr, len(audio))

Governance, privacy & consent

• Informed consent. Every contributor accepts versioned Terms & Privacy (v1 — June 2026) before contributing, on web and in the WhatsApp flow.
• No PII in the dataset. Speakers are referenced only by pseudonymous IDs (SPK01…). No names, phone numbers, or direct identifiers are included.
• Encryption & security. Data in transit over HTTPS; database not publicly exposed; role-restricted access with audit logging; daily encrypted backups.
• Data Protection Act [Chapter 12:07]. Lawful basis = consent; data minimisation; right to withdraw before publication. Full compliance matrix available on request.
• Tiered licensing. Derived transcripts + metadata are open (CC BY 4.0); in the full corpus, raw voice audio is controlled-access. This sample is a consented public demonstration.

Quality & honesty notes

• This is an early, un-incentivised validated slice — deliberately small. The Inzwi thesis is that a proven, governed pipeline + contributor incentives scales this to hours across dialects and into Ndebele (nd).
• Curation: 1 record was excluded from the raw validated set because its prompt text was in English, not Shona — surfaced while building this card. It's a reminder that language-of-prompt checks belong in validation, and a fix is on the Inzwi roadmap.
• Clips range from single words (e.g. Sadza, Nyama, Mbeu) to full sentences, giving both lexical and connected-speech coverage.

Citation

Basarokwe, N. et al. Inzwi: A Community-Driven Speech Corpus for Zimbabwean Indigenous Languages. Sample release, 2026. https://inzwi.app