fix(tts): #91 — rate-floor lift R experiment (sp_it_a_0001)

[shaypal5]

Hypothesis (#91 R)

PR #90's symmetric pitch cap recovered sp_it_a_0001 WER from 0.322 → 0.129, but left a residual gap to the 04-15 baseline (0.056). #89 falsified pitch as the residual driver. R is the next single-knob lever: lift _EFFECTIVE_RATE_MIN from 0.85 → 0.95, on the hypothesis that VIC's I4/I5 slowdown — currently flooring at 0.85 — is what still trips Whisper's silence-detection heuristic.

The risk this PR explicitly traded against: VIC slowdown at high intensity is a deliberate distress cue. Flooring rate at 0.95 may flatten it. The native-speaker listening test was the merge gate — Tier-3 ASR alone is not sufficient.

What this PR ships

A single-constant change plus an anchor test:

file	change
synthbanshee/tts/renderer.py	_EFFECTIVE_RATE_MIN = 0.85 → 0.95. Comment updated to anchor the new value to #91's R rationale. Pitch cap (_EFFECTIVE_PITCH_*) and rate ceiling (_EFFECTIVE_RATE_MAX) untouched.
tests/unit/test_effective_prosody_cap.py	Add test_rate_floor_anchored_to_0_95_for_issue_91 — a literal-value anchor so an inadvertent revert is caught loudly. Existing rate-floor assertions reference the imported _EFFECTIVE_RATE_MIN symbol and auto-track the new value.

Diff: 2 files, +16 / -1.

Tier-3 ASR sanity (local)

sp_it_a_0001 re-rendered at seed 1101 into data/m17_pr89_R_rate_floor/ (5 VIC turns re-rendered against Azure due to new floor; AGG and pitch SSML unchanged → cache hits). Whisper-large-v3 + jiwer over four variants:

variant	dur	WER	length_ratio	hyp / ref
04-15 baseline (target)	155.9s	0.056	1.013	236 / 233
PR #90 reference (cap on)	149.1s	0.129	0.910	212 / 233
B falsification (per-role pitch cap)	149.1s	0.129	0.910	212 / 233
R (rate floor 0.95) — this PR	143.8s	0.052	1.009	235 / 233

WER pass criterion (≤ 0.10) cleared with margin; R reaches WER 0.052, equivalent to the 04-15 baseline. Length-ratio fully recovered to 1.009 (vs PR #90's 0.910).

Cap-clamp breakdown (this render)

19 cap activations across the scene. Rate-floor clamps are the new behaviour from this PR:

turn	role	I	dim	pre	post
04	VIC	2	rate floor	0.912	0.950
06	VIC	2	rate floor	0.901	0.950
08	VIC	3	rate floor	0.878	0.950
10	VIC	3	rate floor	0.846	0.950
12	VIC	3	rate floor	0.819	0.950
14	VIC	4	rate floor	0.833	0.950
14	VIC	4	pitch ceil	+2.29	+2.00
16	VIC	3	rate floor	0.819	0.950

Plus the unchanged PR #90 ceiling clamps on AGG turns (rate 1.20, pitch +2.00) — 11 events on AGG at I3-I5.

Listening test — completed (2026-05-07)

Native-speaker A/B by @shaypal5 between PR #90 reference and R candidate, focused on VIC at I3–I4.

Verdict: parity, not flattening. Both renders sound the same. In both, VIC at I3–I5 does not sound distressed — only the +2.0 st pitch ceiling is audible as an intensity cue, while the rate movement (whether floored at 0.85 or 0.95) is below perceptual threshold.

Verbatim: "both sound the same, which in both cases doesn't sound very distressed, just a robot whose pitch is a bit higher."

What that means for R

R does not regress vs the May-3-calibrated PR #90 cap — naturalness parity holds. The "VIC sounds distressed" bar wasn't being met by main either; that's a deeper TTS issue (rate + pitch knobs are not the right levers for distress in Azure he-IL voices) which is now tracked separately as #97.

Reframed merge rationale: R is a strict WER win at zero perceptual cost. Decoupling the WER fix (this PR) from the distress investigation (#97) is the right call — the two need different levers and shouldn't be gated on each other.

What R does NOT fix

• VIC distress cue at I3–I5 — see TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97 (TTS distress cue absent: rate + pitch are not sufficient signal). Out of scope here.
• AGG aggression at I3–I5 — not exercised by this listening test; may or may not have the same gap.

Test plan

• pytest tests/unit -q — 1688 passed locally (1687 pre-PR + 1 new anchor test).
• ruff check synthbanshee/tts/renderer.py tests/unit/test_effective_prosody_cap.py — clean.
• mypy synthbanshee/tts/renderer.py tests/unit/test_effective_prosody_cap.py — clean.
• Tier-3 ASR sanity (local) — see four-way table above; R clears the 0.10 WER bar and matches the 04-15 baseline. Length-ratio 1.009, no silence-detector trip.
• Native-speaker listening test (paired A/B) — completed 2026-05-07. Verdict: parity vs PR fix(tts): #87 partial — effective-prosody cap addresses Whisper backdoor + helium range #90 reference; no naturalness regression. Distress-cue absence is a pre-existing problem tracked in TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97.

References

• Closes fix(tts): #87 follow-up — test rate-floor lift to address residual sp_it WER gap (R) #91.
• Surfaces TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97 — distress cue at I3–I5 is not addressable through the cap layer's prosody knobs; a different lever (style, breathiness, disfluency, alt voice model) is needed. PR fix(tts): #91 — rate-floor lift R experiment (sp_it_a_0001) #95 banks the WER win without blocking on that work.
• Builds on fix(tts): #87 partial — effective-prosody cap addresses Whisper backdoor + helium range #90 (effective-prosody cap; pitch ceiling + rate ceiling retained from there).
• fix(tts): #87 follow-up — close residual sp_it WER gap (0.129) toward baseline (0.056) without sacrificing M15 calibration #89 (B falsification) — closed on the basis that pitch is not the residual driver; this PR confirms rate is.
• investigate(tts): #83 residual — Whisper WER regression on high-intensity (I3+) Tier A clips #87 — original Whisper WER regression. PR fix(tts): #91 — rate-floor lift R experiment (sp_it_a_0001) #95 fully closes the WER gap on sp_it_a_0001 (0.052 ≈ baseline 0.056).
• CLAUDE.md "ASR sanity check policy" — local-only Tier-3 run completed; paired listening test completed.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR adjusts the TTS “effective prosody” runtime cap to support issue #91’s R experiment, raising the effective rate floor to reduce Whisper WER regressions observed on high-intensity turns (notably sp_it_a_0001) while keeping the pitch caps and rate ceiling unchanged.

Changes:

• Lift _EFFECTIVE_RATE_MIN in synthbanshee/tts/renderer.py from 0.85 to 0.95, with updated in-code rationale tying the change to #91.
• Add a unit-test “anchor” assertion to loudly catch accidental reverts of the new floor value.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
synthbanshee/tts/renderer.py	Raises the effective rate floor constant and documents the #91 rationale/risks inline.
tests/unit/test_effective_prosody_cap.py	Adds a literal-value anchor test ensuring the new floor stays at 0.95.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

pr-agent-context report:

This is a refreshed snapshot of the current PR state.

This run includes a patch coverage gap on PR #95 in repository https://github.com/DataHackIL/SynthBanshee

Address the patch coverage gaps below, then push all of these changes in a single commit.

# Patch coverage

Patch test coverage is 50%; please raise it to 100%. These are the uncovered code lines:
- synthbanshee/tts/renderer.py: 70

Run metadata:

Tool ref: v4
Tool version: 4.0.21
Trigger: schedule
Workflow run: 25480717421 attempt 1
Comment timestamp: 2026-05-07T06:52:07.628445+00:00
PR head commit: b398b4ff68a7ac08d8912aea2c83c7619fc11d65