spike(tts): #97 — Azure express-as fearful style A/B at I4–I5 on sp_it_a_0001

[shaypal5]

TL;DR — negative result, do not merge

A/B-tested Azure <mstts:express-as style="fearful"> on a VIC@I=4 turn in sp_it_a_0001. The two cached audio bytes are byte-identical (f495bf6a…) despite distinct SSML inputs — Azure's he-IL-HilaNeural voice silently ignores the express-as wrapper for this voice + prosody combination. M14's defensive supports_style_tags=False default was empirically correct.

Recommendation: close this PR without merging. Pursue candidate #2 in #97 (disfluency density at high intensity) under the same one-knob discipline.

What changed

File	Change
synthbanshee/tts/renderer.py	Added env var SYNTHBANSHEE_VIC_STYLE_OVERRIDE that, when set AND the turn is VIC at intensity ≥ 4, overrides style_entry.style AND forces supports_style_tags=True per call. The provider's global capability is left at M14's False default. Cache key changes only for the affected turn.
tests/unit/test_vic_style_override.py	New file. Parameterized coverage of the four-condition gate (env-set + role + intensity + provider-default-unchanged) plus multiple style values, so the next style spike needs no code changes.
tests/fixtures/issue_97_fearful_ssml.json	Direct byte-identity evidence: SSML SHAs + cache file pointers + audio SHAs for the General and fearful renders of VIC@I=4 turn 14. Hebrew utterance text redacted per CLAUDE.md.

Out of scope, intentionally:

• No prosody / cap / RMS-gain changes.
• No speaker-YAML edits.
• No SSML-builder changes.
• supports_style_tags global default unchanged (override forces True per call).

Scope-narrowing rationale

Issue says "VIC at I3–I5 does not sound distressed." Override gate is I ≥ 4. Reasoning: the negative result generalises downward — if Azure ignores the express-as wrapper at I=4, it ignores it at I=3 too. Lowering the gate would burn 4× more Azure calls re-rendering I=3 turns for the same finding.

Test plan

• pytest tests/unit/ — 1698 passed (was 1688; +10 new gate tests).
• Ruff + mypy clean on touched files.
• Baseline render of sp_it_a_0001 (env unset): all 17 turns hit cache → $0 Azure.
• Fearful render (SYNTHBANSHEE_VIC_STYLE_OVERRIDE=fearful): only VIC@I=4 turn re-rendered → ~1¢ Azure.
• Verified at the SSML layer that the override fired and the cache key changed.
• Tier-3 ASR sanity (local) — see below.

Tier-3 ASR sanity (local)

$ synthbanshee qa-report data/spikes/97_express_as_fearful/fearful --asr
QA PASSED — failure rate 0.0% ≤ 2.0%
asr_warnings: {}     # 0 clips with length_ratio < 0.85

ASR re-run is mechanically a no-op given byte-identity (the audio is the same as the General-style baseline, which has known WER ≈ 0.052 from PR #95). Included only because CLAUDE.md mandates it for any synthbanshee/tts/ touch.

Direct byte-identity evidence

tests/fixtures/issue_97_fearful_ssml.json records, for the same VIC@I=4 turn (same Hebrew text, same prosody, same voice):

	General	Fearful
xmlns:mstts declared	no	yes
<mstts:express-as> wrapper	no	yes
SSML SHA-256	9880c52f…	18aa1ae1…
Cache file (real Azure render)	9880c52f….wav	18aa1ae1….wav
Audio bytes	309 934	309 934
Audio SHA-256	f495bf6a…	f495bf6a…

Two distinct SSML strings → two distinct cache keys → two real Azure synthesis events → byte-identical audio output. No listening-test interpretation needed.

What this falsifies — and what it doesn't

• Falsified: candidate feat: Phase 0 — pipeline foundation (config schema, TTS, preprocessing, labels, validator) #1 with style="fearful" on he-IL-HilaNeural for this prosody combination.
• Not yet falsified: the other styles in the family (terrified, whispering, shouting, sad). Strong prior they're also no-ops (same voice model), but unverified. One more spike (~1¢ + listening test) would close the family.

Recommendation

• Close this PR without merging. The override mechanism is clean enough to keep on a branch as a reproducer, but landing it as production code is unjustified for a knob we know is decorative on he-IL.
• Next spike: candidate ci: integrate pr-agent-context with coverage XML and refresh flow #2 — disfluency density at high intensity. DisfluencyProfile plumbing already exists; same one-knob discipline.
• A parallel write-up is on #97 so the finding survives this PR's disposition.

Reproducer

git fetch && git checkout spike/97-express-as-fearful-i4-i5
synthbanshee generate -c configs/scenes/she_proves/sp_it_a_0001.yaml -o /tmp/baseline -v
SYNTHBANSHEE_VIC_STYLE_OVERRIDE=fearful synthbanshee generate -c configs/scenes/she_proves/sp_it_a_0001.yaml -o /tmp/fearful -v
cmp /tmp/baseline/agg_m_30-45_001/sp_it_a_0001_00.wav /tmp/fearful/agg_m_30-45_001/sp_it_a_0001_00.wav   # exit 0

Refs

• Refs investigate(tts): #83 residual — Whisper WER regression on high-intensity (I3+) Tier A clips #87 (already closed in spirit by PR fix(tts): #91 — rate-floor lift R experiment (sp_it_a_0001) #95; closure comment left).
• Refs TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97 (falsifies candidate feat: Phase 0 — pipeline foundation (config schema, TTS, preprocessing, labels, validator) #1; issue stays open pending candidate ci: integrate pr-agent-context with coverage XML and refresh flow #2).

[Copilot]

Pull request overview

This PR introduces a targeted spike to A/B-test Azure SSML <mstts:express-as> styles for VIC turns at high intensity (I≥4) by (1) enabling style-tag capabilities for Azure and (2) adding an env-var-driven override in the renderer, with a corresponding unit test update. The PR description reports a negative result (Azure he-IL voice ignores express-as), which makes the current capability flip and spike plumbing unsuitable to merge as-is.

Changes:

• Flip Azure provider supports_style_tags from False to True.
• Add SYNTHBANSHEE_SPIKE_97_VIC_STYLE env-var override in TTSRenderer.render_utterance() for VIC at intensity ≥ 4.
• Update unit test expectations to match the new Azure capability value.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
synthbanshee/tts/azure_provider.py	Advertises Azure backend style-tag support by toggling supports_style_tags to True.
synthbanshee/tts/renderer.py	Adds spike-specific env-var logic to override per-turn style before SSML generation.
tests/unit/test_google_provider.py	Updates TestProviderCapabilities.test_azure_capabilities to assert supports_style_tags=True and adds rationale comments.

---

💡 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 #98 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 62.50%; please raise it to 100%. These are the uncovered code lines:
- synthbanshee/tts/renderer.py: 80, 94, 95, 96, 97, 103

Run metadata:

Tool ref: v4
Tool version: 4.0.21
Trigger: workflow dispatch
Workflow run: 25574244529 attempt 1
Comment timestamp: 2026-05-08T19:08:03.765203+00:00
PR head commit: fbd83cee5cb992c92428e8a0d5267af01f23cdef

[shaypal5]

Closing without merge — spike concluded with a negative result.

Disposition

• fearful on he-IL-HilaNeural confirmed as a no-op via byte-identity at the cache layer (see tests/fixtures/issue_97_fearful_ssml.json on this branch).
• The override mechanism in synthbanshee/tts/renderer.py was the spike's reproducer; it evaporates with this PR.
• Branch preserved as the reproducer — git checkout spike/97-express-as-fearful-i4-i5 to re-run the A/B.

Successor work

• Falsification of the candidate-feat: Phase 0 — pipeline foundation (config schema, TTS, preprocessing, labels, validator) #1 family (one more style spike, e.g. terrified) — optional confirmation that all he-IL express-as styles are no-ops.
• Otherwise: candidate ci: integrate pr-agent-context with coverage XML and refresh flow #2 in TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97 — disfluency density at high intensity.

Cross-references

• Refs investigate(tts): #83 residual — Whisper WER regression on high-intensity (I3+) Tier A clips #87 (closed in spirit by PR fix(tts): #91 — rate-floor lift R experiment (sp_it_a_0001) #95).
• Refs TTS distress cue absent at I3–I5: rate + pitch are not sufficient signal #97 (candidate feat: Phase 0 — pipeline foundation (config schema, TTS, preprocessing, labels, validator) #1 partially falsified; issue stays open).