investigate(tts): same scene config now yields 63% longer audio (121s → 198s) — confirm new duration regime is intentional

[shaypal5]

Discovered while re-rendering sp_neu_a_0001 from its scene config on current main for #78.

Evidence

Same scene config (configs/scenes/she_proves/sp_neu_a_0001.yaml, 12 turns, intensity arc [1,1,1,2,1], AGG_M_30-45_001 + VIC_F_25-40_002), rendered before and after the prosody changes that landed in late April / early May:

Date	Code path	Duration	Vs scene target_duration_minutes: 3.0 (180 s)	Vs She-Proves window (3–6 min)
2026-04-15	pre-M15 / pre-#70 / pre-#68	121.0 s	33% under target	below lower bound
2026-05-05	current main	197.8 s	10% over target	within range

The new duration is actually closer to the spec target than the old one. But the +63% delta is large enough that downstream code that estimated wall-cost or filename-budget on Tier A clips against the old duration regime is now off.

Likely cause (not yet confirmed)

Cumulative effect of recent prosody work, in rough order of suspected magnitude:

• feat(m15): SSML prosody tuning with research-validated Hebrew parameters #51 (M15) — SSML prosody tuning with research-validated Hebrew parameters. Changes rate multipliers in speaker style_map.
• fix(tts): insert inter-word <break> tags to prevent Hebrew word merging #70 — inter-word <break> tags to prevent Hebrew word merging. Each break adds a small pause; for ~100 Hebrew words per turn over 12 turns, even a 50 ms median break adds ~60 s to the clip.
• feat(m14): fix muffled audio, click artifacts, and voice identity shifts #48 (M14) — fix muffled audio / removed 7500 Hz LPF. Should not affect duration directly, but the M14 PR also changed the edge-fade behaviour; check whether anything there affects perceived/measured speech end.

A controlled bisect on the same scene config (cached LLM script + cached SSML where possible, only the rendering path changing) will identify the dominant contributor cheaply.

Why this is not a #78 dependency

#78 is purely about loudness (peak / RMS). Duration is independent. But both regressions came in the same window of PRs and surface together when re-rendering Tier A clips, so it's worth tracking the duration delta separately so the loudness fix doesn't accidentally take ownership of "why are clips longer now."

Decision needed

Two questions:

1. Is the new duration regime intentional? Per CLAUDE.md, She-Proves clips should be 3–6 min. 198 s = 3.3 min satisfies the lower bound; 121 s did not. If the team agrees the new duration is correct, this issue closes as "investigate, document, no code change."
2. Does anything downstream budget on duration? E.g. label-generator phase boundaries, augmentation event placement, M17 evaluation runtime estimates. If yes, those budgets need to be re-derived against current TTS output.

Reproduction

.venv/bin/synthbanshee generate -c configs/scenes/she_proves/sp_neu_a_0001.yaml \
    -o /tmp/duration_repro -p she_proves
.venv/bin/python -c \"import soundfile as sf; print(sf.info('/tmp/duration_repro/agg_m_30-45_001/sp_neu_a_0001_00.wav'))\"

Old reference at data/m2a_wettest/agg_m_30-45_001/sp_neu_a_0001_00.wav is 121.0 s.

References

• Discovered alongside loudness regression in bug(tts): rendered clip loudness regressed by ~6 dB between 2026-04-15 and 2026-05-05 #78.
• Suspect commits: feat(m15): SSML prosody tuning with research-validated Hebrew parameters #51 (M15), fix(tts): insert inter-word <break> tags to prevent Hebrew word merging #70, feat(m14): fix muffled audio, click artifacts, and voice identity shifts #48 (M14).