perf(desktop): stream Pocket TTS decoder chunks

[johnmatthewtennant]

Context

Pocket TTS waits for a complete synthesis unit before queuing audio even though its decoder produces usable PCM blocks during generation. Long agent replies therefore begin later than necessary.

The callback is cumulative and may repeat a length between internal model chunks, so playback must queue only each new suffix and reject decreasing lengths.

Summary

Queue Pocket TTS decoder output as it arrives while the resident engine continues synthesizing. Playback remains ordered and cancellable, including during silent decoder gaps.

Changes

• Expose callback-driven Pocket synthesis while preserving the buffered API.
• Convert cumulative PCM into ordered suffixes and fail closed on callback contract violations.
• Keep only the final tail needed for the utterance fade, then queue audio while synthesis continues.
• Keep the first sentence as a playback boundary while handling 50-token model splits internally, without extra playback pauses.
• Keep synthesis activity armed through final queueing so PTT, barge-in, preview completion, voice changes, and shutdown observe the correct state.
• Cover ordering, repeated callback lengths, cancellation, quiet onsets, underrun restart, queue errors, silent-gap microphone gating, and playback grouping.

Related issue

None found.

Testing

Manual validation in the combined daily-driver build confirmed that long replies begin playing before synthesis completes, decoder boundaries have no repeated or omitted audio, PTT and barge-in cancel during active playback and silent decoder gaps, and voice Preview completes.

For the same long input, warm Pocket TTS produced its first usable callback at 1020.0 ms and completed synthesis at 3220.9 ms. The output was 16.640 seconds of audio at 0.194 RTF. Buffered and streaming assembly produced byte-identical PCM, and callback cancellation returned in 0.0 ms.

Screenshots

Not applicable. This changes the native audio pipeline and has no visual UI change.

Reviewer-reproducible examples

From a fresh checkout with the Pocket voice model installed:

. ./bin/activate-hermit
just dev

1. In Voice settings, enable Agent text to speech and select a Pocket voice.
2. Join an agent huddle and request a long, multi-sentence response.
3. Confirm speech begins before the response finishes synthesizing and remains continuous at sentence and model-token boundaries.
4. Hold PTT or speak once during audible playback and once during a silent decoder gap. Confirm the remaining speech stops immediately and microphone input continues.
5. Run Voice Preview and confirm it plays to completion.

Observed on the PR build: speech began during synthesis, no repeated, omitted, gapped, clicked, or clipped audio was heard, both cancellation paths stopped the remaining speech, and Preview completed.