forked from akai/readest
Huang Xin
42f9b8fe3c
* feat(tts): add PCM speech-bounds detection for sentence audio trimming Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): add WSOLA time-stretch for pitch-preserved playback rate Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): add sentence duration store with per-voice speaking-rate calibration Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * refactor(tts): serve edge audio as ArrayBuffer with in-flight fetch dedup Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): add WebAudioPlayer with gapless chunk scheduling and backpressure Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): play edge TTS through gapless Web Audio pipeline Replaces the per-sentence audio element with trimmed, time-stretched buffers scheduled on the shared AudioContext. Marks dispatch at audible time so schedule-ahead cannot run foliate's cursor past the voice; a decode failure or missing audio skips the chunk instead of wedging the session; pause and resume ride context suspend and resume with no iOS rewind hack; the object-URL cache is gone. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): add section timeline with measured and estimated sentence durations Includes the foliate-js submodule bump for the getSentences export (fork branch feat/tts-get-sentences; fork PR must merge before this lands so the pinned SHA resolves). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): expose section playback position and sentence-snapped seeking Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): surface playback position and seek in the media session Position state is clamped, never skipped, so the lock-screen scrubber stays live when estimates overshoot; seekto units map per backend (native ms, web seconds). The AudioContext warms up in the tts-speak gesture path before any await, and the silent keep-alive element now runs on all platforms so desktop hardware media keys survive the removal of the per-sentence audio element. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): add seekable chapter progress bar to the TTS panel The scrubber joins the transport cluster with a thin range-xs track and flanking tabular time labels so it cannot be misgrabbed for the chunky rate slider (which persists a global setting). States: reserved disabled slot until the lazy timeline lands, persists across chapter transitions, optimistic thumb with failure toast, monotonic position, tilde-prefixed estimated totals, sentence-event updates under e-ink. The popup grows only when a timeline-capable client is active. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: record deferred TTS listening-engine follow-ups Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: record background TTS decoupling design decisions Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(tts): slim the panel scrubber to a native track with remaining time Match the footer Jump to Location slider (plain native range: thin track, small thumb) instead of the chunky daisyUI pill, show remaining time with a minus prefix on the right, and drop the This chapter caption. Popup height shrinks accordingly. Verified live in Chrome. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * chore: pin foliate-js to merged main with getSentences export Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(tts): catch autoplay rejection from the keep-alive element Running the silent keep-alive on all platforms exposed an un-awaited play() that headless Chromium rejects without a user gesture, failing CI on unhandled rejections while every test passed. The keep-alive is best-effort; the production path is gesture-qualified. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1.8 KiB
1.8 KiB
TODOS
Deferred items from the Edge TTS Web Audio plan review (/autoplan, 2026-07-04).
Each was explicitly deferred, not forgotten — see the Decision Audit Trail in
.agents/plans/2026-07-03-edge-tts-webaudio.md.
TTS listening engine follow-ups
- Cross-section gapless playback: preload and schedule the next section's first sentence so chapter boundaries are as seamless as sentence boundaries. (M)
- Buffer-ahead indicator in the TTS scrubber (show the prefetched region, YouTube-style). (S)
- Lock-screen ±10s seek offsets in addition to prev/next sentence. (S)
- Persist measured sentence durations per book so a reopened chapter starts with an exact timeline instead of estimates. (S)
- Worker offload for decode + WSOLA if device profiling shows main-thread jank on low-end Android. (S)
- Sticky TTSBar scrubber (panel + lock screen only in the first release; recorded as deliberate in decision #24). (S)
- Provider-agnostic voice source hedge: local neural TTS (e.g. Piper/Kokoro WASM) plugging into WebAudioPlayer/SectionTimeline — the engine is designed for this; see "Strategic framing" in the plan. (L)
- Background chapter prefetch (convert timeline estimates to exact durations ahead of playback). (M)
- Background TTS: decouple session ownership from the reader view via an app-level TTSSessionManager so closing the book keeps TTS playing (headless text supply via section.createDocument(), CFI re-anchoring for highlights on reattach, library now-playing pill). Decided matrix: close book = keep playing; reopen same book = seamless reattach (adopt session, redispatchPosition, lazy doc swap at next section); open a DIFFERENT book = TTS stops; explicit stop / sleep timer = stops. (M)