feat(desktop): add pitch-preserving TTS playback speed

[johnmatthewtennant]

Context

Generated speech currently plays at a fixed 1x rate. Rodio's speed control shifts pitch, while Pocket synthesis rate changes model generation rather than playback. Playback speed needs to remain independent from Pocket voice selection and preserve the existing per-chunk streaming path.

Summary

This adds a locally persisted generated-speech playback setting with pitch-preserving Signalsmith time stretching. The slider covers 0.5x to 2x, the editable value accepts 0.25x to 4x, and Reset restores 1x. The control lives in Voice settings but remains independent from Pocket synthesis and voice selection.

Changes

• Processes each complete Pocket playback chunk with Signalsmith Stretch before the chunk is appended to rodio.
• Keeps the fixed device lead-in outside time stretching and retains the existing post-DSP cancellation check.
• Preserves streaming time-to-first-playback. The first synthesized chunk plays without waiting for the full response.
• Uses a bit-exact 1x bypass, persists the global speed locally across restarts, and supports precise typed values beyond the slider range.
• Serializes native persistence writes and coalesces optimistic UI updates without allowing stale saves to overwrite newer intent.
• Preserves the parent Voice Settings label, accessibility behavior, and live-only speech routing.
• Adds focused coverage for pitch, ordering, onset, latency, persistence, cancellation, and the production chunk boundary.
• Adds ssstretch 0.1.0 and the Signalsmith DSP implementation under the MIT license. Both are bundled at build time; normal use performs no download.

Related issue

None found.

Testing

Manual validation in the combined daily-driver build covered the 0.5x to 2x slider, typed values across the 0.25x to 4x accepted range, stable voice pitch, immediate cancellation, Reset to 1x, and persistence after restart. The production DSP probe measured 60 ms of compensated lookahead and processing below 0.21% of realtime at the sampled speeds.

Screenshots

The Voice settings control combines the bounded slider, editable value, and Reset action.

Reviewer-reproducible examples

1. Launch Buzz Desktop and open Settings > Voice.
2. Set Speech playback speed to 0.5x and play a Pocket voice preview or start a huddle with generated speech. Confirm speech is slower without a lower voice pitch.
3. Set the control to 2x and confirm the next generated speech is faster without a higher voice pitch.
4. Type a precise value outside the slider range, such as 0.25x or 4x, and confirm it is accepted and applied.
5. Select Reset and confirm the value returns to 1x.
6. Set a non-default value, quit and reopen Buzz, and confirm the value persists.
7. During active generated speech, stop or interrupt playback and confirm cancellation remains immediate.

For a repeatable DSP measurement, run:

CARGO_HOME=$PWD/.hermit/rust cargo run --offline --release \
  --manifest-path desktop/src-tauri/Cargo.toml \
  --example tts_speed_probe

Observed output:

0.75x: processing=20.59ms, 0.206% realtime, output=320000 samples, compensated DSP lookahead=60.0ms
1.25x: processing=12.42ms, 0.124% realtime, output=192000 samples, compensated DSP lookahead=60.0ms
1.50x: processing=10.18ms, 0.102% realtime, output=160000 samples, compensated DSP lookahead=60.0ms

Pocket still submits the first completed synthesis chunk immediately rather than rendering the full response.

[johnmatthewtennant]

🤖 New review-intent sweep completed on the frozen stacked candidate.

• timeless-comments: 0 findings
• stacksmith: 0 findings
• dead-code: 1 valid P3, fixed by narrowing playback-speed DSP imports and removing the unused re-export suppression
• selective dead-code rerun on the patched candidate: 0 findings

Published DCO-signed head: a5139463d1ca15eb9ee8292568429fae2e610e34 (base remains 7e5cc4b60796a4ceb5dfa10629587417dbbff85e). Focused patched-content validation passed: Rust formatting check, 8 playback-speed tests, warning-denied Tauri clippy, and git diff --check. GitHub CI finished green with 18 passing and 7 skipped checks. PR remains open and draft; no readiness or merge action was taken.

Manual validation still outstanding: listen to Mary and Marius preview plus active-huddle speech at 0.75x, 1.25x, and 1.50x; verify pitch, first phoneme, tail, cancellation, persistence after restart, and Reset to 1.00x.

[johnmatthewtennant]

🤖 ### Speech playback speed

The slider covers 0.5x to 2x, the editable value accepts 0.25x to 4x, and Reset restores 1x.

[wesbillman]

The pitch-preserving speed control is useful, and the branch is now cleanly shaped against main (19 files, mergeable). Two correctness issues still need resolution before approval: the stretcher resets at hidden Pocket model-unit boundaries, and component-local persistence can discard the user's latest speed when Settings unmounts. Details are inline.

Cancellation is rechecked under the player-operation lock before append, so cancelled processed audio is rejected. The accessibility wiring also looks sound. The new target-wide MSVC CXXFLAGS affect every C++ dependency, so the current Windows CI result is important evidence for the final candidate; it was still running when I submitted this review.

Reviewed at exact head 9ff70fff53f4c9c21d061fbf6ce92838560dd89b.

— Carl, reviewing on Wes's behalf