fix(voice): improve Pocket playback speed quality

[johnmatthewtennant]

Summary

Pocket speech at accelerated and reduced speeds on macOS now uses AVAudioUnitTimePitch, preserving natural pitch while avoiding the metallic and choppy artifacts from the previous time-stretching path. Playback at 1x bypasses the effect, selected output-device routing remains supported, and interruption accounting reports only audio that the native player completed.

The lower-quality sstretch dependency and cross-platform fallback are removed. Pocket playback is macOS-only in this final state.

Related issue

None found.

Reviewer-reproducible examples

1. On macOS, select Pocket TTS in Voice settings and choose an accelerated playback speed such as 1.5x or 2x.
2. Start a voice conversation and request a multi-sentence response.
3. Confirm the speech keeps a natural pitch and avoids metallic or choppy artifacts.
4. Switch to 1x and confirm normal-speed playback remains natural. Select a non-default audio output and confirm Pocket playback uses it.

Testing

John manually compared accelerated Pocket playback in a macOS development build and accepted the improved audio quality.

[morgmart]

🤖 Automated code review

REQUEST_CHANGES: one blocking lifecycle regression is present in the exact three-dot comparison. Pocket playback now drops its assistant-speech/capture-protection guard as soon as AVAudioPlayerNode reports its final source buffer completed, removing the existing route-aware output-latency grace; buffered Bluetooth, AirPlay, virtual, and unknown routes can therefore resume capture while speech is still physically audible. I statically reviewed all eight changed files and the complete Pocket playback flow (creation, routing, speed processing, enqueue, progress accounting, completion, interruption, failure, and cleanup), plus accessibility, localization, navigation/consent, async never-completes and races, tests, project laws/design-system applicability, and overlap/duplicate status. No additional supported findings survived the final self-check. Supplied GitHub evidence was inspected: the listed DCO, frontend, clean-room, and platform Rust checks succeeded, while combined status was pending; required checks still govern merge readiness. No local code, tests, builds, package managers, hooks, binaries, or project scripts were run.

Deterministic publication result: 1 blocking and 0 non-blocking finding(s) publishable; 0 duplicate(s) suppressed.

[morgmart]

🤖 Automated code review

REQUEST_CHANGES: one blocking lifecycle regression is present in the exact three-dot comparison. The new finish path has no degraded path if the native pending-buffer completion bookkeeping never reaches zero, so Pocket playback can remain active indefinitely; the replaced implementation explicitly bounded this condition and its regression test was deleted. I statically reviewed all eight changed files and the complete Pocket playback flow (player creation, speed processing, output routing, enqueue, progress/delivery accounting, burst/grace handling, completion, interruption, failure, and cleanup), plus accessibility, localization, navigation/consent, async error/never-completes/race behavior, test honesty, project laws/design-system applicability, and finding overlap. The prior output-latency capture-grace issue is fixed in the current comparison and its resolved thread has a substantive human reply, so it is neither a new finding nor a duplicate suppression. No other supported findings survived the final self-check. Supplied GitHub evidence was inspected: DCO, frontend, clean-room, and platform Rust checks were successful, Transcript virtualization was in progress, and combined status was pending; required checks still govern merge readiness. No local code, tests, builds, package managers, hooks, binaries, or project scripts were run.

Deterministic publication result: 1 blocking and 0 non-blocking finding(s) publishable; 0 duplicate(s) suppressed.

[morgmart]

🤖 Automated code review

APPROVE: no publishable findings remain in the exact three-dot comparison. I statically reviewed all eight changed files and traced every changed Pocket flow: native player creation and ownership, speed processing, output routing, streaming append/flush/finish, one-shot preview/speak, enqueue and delivery accounting, route-aware capture grace, timeout fallback, interruption, stop, failure, and cleanup. The stream finish path now bounds missing native completion bookkeeping and independently preserves route-aware capture grace. A similar never-completes risk remains in the changed one-shot preview/speak drain loop, but it is the same underlying missing-native-completion issue as the resolved protected thread whose substantive human reply states the fix scope; policy therefore suppresses it rather than publishing it again. Final checks covered accessibility and i18n (no UI/string-surface change requiring localization), navigation/consent (not affected), async errors, never-completes, lifecycle and races, test honesty, design-system/project rules, duplicate overlap, and concrete user effects; no other supported issue survived. Supplied GitHub evidence was structurally valid and inspected: DCO, frontend, transcript virtualization, clean-room, and macOS/Windows/Linux Rust checks all completed successfully, while combined status was pending; required checks still govern merge readiness. No local code, tests, builds, package managers, hooks, binaries, or project scripts were run.

Deterministic publication result: 0 blocking and 0 non-blocking finding(s) publishable; 1 duplicate(s) suppressed.

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.

[morgmart]

🤖 Automated code review

APPROVE: no publishable findings remain in the exact three-dot comparison. I statically reviewed all eight changed files and traced every changed Pocket flow: native player creation and ownership, AVAudioUnitTimePitch speed processing, output-device routing, streaming append/flush/finish, one-shot preview/speak, enqueue and delivery accounting, route-aware capture grace, bounded native-completion fallbacks, interruption, explicit stop, output failure, and cleanup. Both previously reported issues are fixed in the current comparison: streaming playback preserves route-aware capture protection after drain, and missing completion bookkeeping is bounded for both streaming and one-shot playback. Final self-checks covered every changed user flow; accessibility and i18n/localization (no changed visual, interactive, or translatable UI surface); navigation and consent guards (not affected); async errors, never-completes, lifecycle, cancellation, and races; test honesty and discriminating coverage; project laws and design-system rules; duplicate/overlapping findings; and concrete evidence plus user effect for blocking severity. No supported concern survived. Supplied GitHub evidence was structurally valid and inspected: DCO, frontend build/tests, transcript virtualization, clean-room install, and macOS/Windows/Linux Rust checks all completed successfully, while combined status was pending; required checks still govern merge readiness. No local code, tests, builds, package managers, hooks, binaries, or project scripts were run.

Deterministic publication result: 0 blocking and 0 non-blocking finding(s) publishable; 0 duplicate(s) suppressed.

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.