Voice Scale — Audio QA Pipeline

Automated quality assurance pipeline for the Voice Scale crowdsourced vegetable-name dataset (English, Tamil, Sinhala). Runs 14 checks across four layers before any clip is accepted into the final dataset.

---

Pipeline overview

Raw WAV clips
     │
     ▼
Layer 1 — Signal Checks        T-01 duration & silence
  (runs on every submission)   T-02 clipping & distortion
                               T-03 SNR
                               T-04 sample rate & format
     │
     ▼
Layer 2 — Transcription        T-05 Whisper (English)
  (catches wrong words)        T-06 Whisper (Tamil)
                               T-07 language ID
                               T-08 duplicate detection
     │
     ▼
Layer 3 — Embedding Analysis   T-09 wav2vec2 extractor  ← must run first
  (catches semantic outliers)  T-10 per-class outlier detector
                               T-11 cross-language consistency
                               T-12 UMAP visualisation
     │
     ▼
Layer 4 — Reporting            T-13 aggregator & HTML dashboard
                               T-14 speaker-level fairness audit

Every task in layers 1–3 writes a CSV with columns clip_id, verdict, reason, score. T-13 merges them all.

---

Task assignments

Task	Title	Owner
T-01	Duration & silence validator	Praveen
T-02	Clipping & distortion detector	Ashini
T-03	SNR estimator	Chanupa
T-04	Sample rate & format validator	Didula
T-05	Whisper transcription runner (English)	Sheron
T-06	Whisper transcription runner (Tamil)	Thila
T-07	Language ID verifier	Kusal
T-08	Duplicate & near-duplicate detector	Lasana
T-09	wav2vec2 embedding extractor	Sahe
T-10	Per-class outlier detector	Teshan
T-11	Cross-language consistency checker	Yohan
T-12	UMAP visualisation generator	Ifaz
T-13	QA results aggregator & dashboard	Thisen
T-14	Speaker-level fairness auditor	Thisen

---

Key dependencies

• T-09 must complete before T-10, T-11, T-12 (they all read its .npy output)
• T-13 and T-14 must run after all layer 1–3 tasks are complete

---

Docs

• tasks/layer1_signal_checks.md
• tasks/layer2_transcription.md
• tasks/layer3_embeddings.md
• tasks/layer4_reporting.md
• COMMIT_CONVENTION.md