Skip to content

Release 0.0.4-beta.1: no lost tail words, offline punctuation, redesigned app pages - #124

Merged
Slacker-LLC merged 3 commits into
mainfrom
codex/beta4-voice
Oct 4, 2026
Merged

Slacker-LLC merged 3 commits into
mainfrom
codex/beta4-voice

Conversation

@Slacker-LLC

@Slacker-LLC Slacker-LLC commented Oct 4, 2026 •

Copy link
Copy Markdown
Owner

Summary

  • Release the microphone the instant the space key is released and decode already-recorded audio to the end. Streaming Paraformer needs is_final to decode its last short chunk, so trailing words were being dropped. The 300 ms post-release capture is removed; media volume returns when capture actually stops.
  • Add the CT-Transformer INT8 offline punctuation model (loaded on demand) and apply "punctuation as space" after punctuation restoration.
  • Batch PCM copies, reuse inference buffers, and report buffer overflow as an error instead of committing a truncated result.
  • Redesign the app home and settings pages (16dp margins, iOS-style switch and pill buttons, skin sub-page, large-font and wide-screen layouts). The keyboard itself is unchanged.
  • Switch colours and layout text sizes live in the design tokens; the trusted voice model catalog and its test are synced with the four-file manifest.

Test plan

  • testDebugUnitTest 383/383 (first run had 3 failures, fixed in this PR), lintDebug
  • scripts/beta4_e2e.py 7/7 on an API 36 x86_64 emulator (real native model, real space touch, real AudioRecord)
  • scripts/core_regression.sh 19/19, scripts/beta3_e2e.py 10/10 (portrait + landscape), scripts/display_matrix_regression.py 11/11
  • Voice failure paths with real touch (hold durations, swipe-up cancel, hide keyboard / HOME mid-recording, 10 rapid cycles, silence, microphone permission revoked and re-granted) 50/50
  • Not covered: real-device microphone accuracy and arm64 hardware (CI release rehearsal builds the signed arm64 APK)

🤖 Generated with Claude Code

Slacker-LLC and others added 3 commits October 4, 2026 15:14
…gned app pages

- Stop the microphone the instant the space key is released and decode the
  already-recorded audio to the end. Streaming Paraformer needs is_final to
  decode its last short chunk; without it trailing words were dropped. The
  300 ms post-release capture is gone and media volume returns when capture
  actually stops.
- Add the CT-Transformer INT8 punctuation model (offline, loaded on demand)
  and apply "punctuation as space" after punctuation restoration.
- Batch PCM copies, reuse inference buffers and report buffer overflow as an
  error instead of committing a truncated result.
- Redesign the app home and settings pages (16dp margins, switch and pill
  button styling, skin sub-page, large-font and wide-screen layouts).
- Keep switch colours and layout text sizes in the design tokens, and sync
  the trusted voice model catalog and its test with the four-file manifest.
- Add a debug-only real-audio-to-editor E2E and scripts/beta4_e2e.py.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… mode

- Home: a three-step setup (progress bar, current step highlighted, bottom
  button performs the current step) that folds into a single "ready" card
  with the try-typing field, microphone status and shortcuts once openIME
  is enabled and selected. App strings are one Chinese set; the partial
  English translation produced half-English pages.
- Preferences: grouped cards with inset hairlines and a neutral icon per
  row instead of one key-cap card per row. Appearance options stack when
  the measured labels do not fit the width and font scale.
- Keyboard settings panel: four quick toggles plus appearance and height on
  the first screen, compact rows below. Tools panel: 56dp icon tiles.
- Accent only on interactive controls and on/current state; switches use
  the accent instead of green. New light and dark app palettes; the
  keyboard keys keep their reference design.
- Key haptics use KEYBOARD_TAP (VIRTUAL_KEY on API 26); gesture thresholds
  use CONTEXT_CLICK instead of LONG_PRESS.
- Remove keyboard skins (accent, radius, opacity, key font size) and
  one-handed mode from settings, storage and the data archive. Old saved
  values and archive fields are ignored.
- Update APP_UI_SPEC, CHANGELOG, tests and the beta4/design E2E scripts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The keyboard settings panel now shows 数字和符号提示 as a quick-toggle
tile instead of a row switch, so the test helper accepts either.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Slacker-LLC
Slacker-LLC merged commit b863fbe into main Oct 4, 2026
7 checks passed
@Slacker-LLC
Slacker-LLC deleted the codex/beta4-voice branch October 4, 2026 09:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant