Dataset

Status: synthetic development corpus. It validates the pipeline but is not human speech and is not human-reviewed.

Contents

Three public-domain US speech excerpts: Gettysburg Address, Second Inaugural Address, FDR First Inaugural Address.

Per speech, 21 items:

Group	Items
Near-perfect mirrors	2 (same speaker; different speaker), with natural-variation jitter
Flaw families × severity 1–3	pace_fast, pace_slow, pause_excess, pause_missing, pitch_flat, energy_quiet (18 items)
Combo	1 item with pacing + pitch + pause flaws

Total: 63 items, about 70 MB.

How it's made

• Baselines: espeak-ng speaks each word separately and the words are joined with controlled gaps (0.06 s inside clauses, 0.40 s after commas, 0.70 s after sentences), so word timings are exact.
• Participants: re-synthesised in the same or a different voice, with jitter on pitch, gaps and speed.
• Flaws (Tier-1): word-aligned edits with exact ground truth: time-stretch, extra or removed pauses, pitch-contour compression via Praat, gain reduction.
• Everything is seeded and deterministic: make dataset reproduces the same items.

Manifest (manifest.jsonl)

Each row: item_id, speech_id, reference_id, audio, transcript, flaw_type, severity, voice, cross_speaker, ground_truth_spans, audio_format, provenance, label_status, label_tolerance_s. items/<id>.json has exact word timings and the flaw plan. references/<id>/reference.json has transcript, provenance and license.

Ground-truth conventions: word-range flaw = first word start to last word end; excess pause = the gap; missing pause = junction ± 0.2 s.

What's still needed for the real submission

1. Real public-domain baseline recordings (avoid CC BY-NC-ND traps) with verified reuse terms for the recording itself.
2. Human re-recordings of the same texts across a quality spectrum.
3. Human review of alignment and flaw boundaries, with a recorded tolerance. Don't claim millisecond accuracy unless the annotation process supports it.
4. A public home for the data (GitHub or public Drive link in the README).