Roadmap

Internal deadline 14 Oct 2026. The event page lists 12:15 AM IST, 15 Oct 2026, so don't plan work to the last minute. Feature freeze: 12 Oct.

Dates	Focus	Exit criteria	State
Oct 1–2	Requirements and design	One sample item and an agreed data contract	✅
Oct 3–5	Thin end-to-end slice	One reference plus one flawed mirror gives a visible result	✅ (backend) · 🟡 (UI)
Oct 6–8	Dataset and feature pipeline	Every item has provenance, transcript, labels, metadata; failure cases logged	🟡 synthetic done, real data pending
Oct 9–10	Integration and evaluation	Full flow from a fresh environment, no hand-edited output	⬜
Oct 11–12	Hardening and freeze	Pinned deps, repeatability checks, saved evaluation	⬜
Oct 13	Documentation and video	Tech doc (max 6 pages), README, 3–10 min video with subtitles	⬜
Oct 14	Submission and buffer	All links verified from a clean session	⬜

Next up (priority order)

1. Dashboard connected to the API (upload, playback, transcript, overlays, scores, explanation cards, error and low-confidence states).
2. Real public-domain baseline recordings with checked reuse terms.
3. Human re-recordings and human-reviewed labels.
4. Validate WhisperX forced alignment on real audio; re-run evaluation.
5. Noise-robust speech/pause detection and a stricter SNR rule.

Later / stretch

• More speeches and speakers; multiple references per text.
• Additional rubric dimensions (e.g. articulation clarity).
• Confidence intervals on metrics once the dataset is larger.

Not planned

• LLM-based scoring. Determinism and traceability are core goals.
• Claims about intent, emotion or universal correctness.

Judging weights (guide for prioritisation)

Criterion	Weight
Data engineering and stress testing	30%
Causal explainability and temporal grounding	25%
Feature extraction	20%
Visualization and dashboard	15%
Reproducibility and code quality	10%