Skip to content

feat: voice quality ports (Skip cleanup, mic test, filler removal, Silero VAD) - #87

Merged
softmaxe merged 16 commits into
mainfrom
feat/voice-quality-ports
Nov 12, 2026
Merged

softmaxe merged 16 commits into
mainfrom
feat/voice-quality-ports

Conversation

@softmaxe

Copy link
Copy Markdown
Owner

Summary

Four dictation-quality features, plus pill polish tuned against real server timings.

 dictation
   record
+    Silero VAD (embedded ONNX) drops silence before upload
   transcribe
   cleanup
+    Skip: deliver the raw transcript now, keep the transcription
+  remove filler words (CJK-aware, keeps quotes, code, casing, commas)
   deliver

 settings
+  Microphone: record a private ≤10 s test clip and play it back
+  Transcription options get their own section

Pill pacing during cleanup, measured against the live ASR + cleanup servers:

cleanup running
  < 400 ms   keep "Transcribing" (no stage flash)
  ≥ 400 ms   swap label to "Polishing"
  ≥ 1.5 s    slide in the Skip chip
Audio Transcribe Cleanup Total
2.5 s en ~130 ms ~170 ms ~300 ms
10 s zh ~280 ms ~350–440 ms ~650 ms
32 s zh ~750 ms ~1000 ms ~1.8 s

New dev hook: WHISPER_DEV_SCENE=pill:flow loops one whole dictation; WHISPER_DEV_FLOW_MS=<transcribe>,<cleanup> replays any server speed (see docs/testing.md).

Evidence

  • Before: at measured speeds the pill flashed "Polishing" and an unreachable Skip for 170–440 ms on every short dictation.
    After: short dictations go recording → Transcribing → check; Polishing appears only past 400 ms, Skip only past 1.5 s.
  • scripts/check.sh: 648 passed, 0 failed, 14 ignored (hardware).
  • New tests include fast_cleanup_never_shows_its_label_or_skip, swapped_layers_never_overlap_at_full_strength, processing_layout_centers_the_label_and_docks_the_skip_chip, plus VAD fixtures (tests/fixtures/vad-*.json|wav) and filler-word cases.
  • Not yet verified: a live microphone dictation in the signed app; hover and Reduce Motion on the Skip chip.

Merge Danger

Door: two-way

Revert restores the previous pipeline; settings additions are additive.

Blast Radius: dictation

VAD and filler removal change what text every dictation delivers; a mis-trimmed segment or over-eager filler edit would show up as missing words. Adds a ~1.3 MB ONNX model to the binary.

Skip silence and Remove filler words are not language choices. The
microphone test's playback track now fills the strip so Close sits at
its trailing edge.
Removed fillers now go as one group with their own punctuation, and the
seam keeps only what the surviving clause needs: a terminal closes the
clause before it, a comma that only introduced the filler is dropped.
Inline code, quoted and bracketed mentions are left alone.
A comma before a filler stays even at a line end, so letter salutations
survive. Recapitalization skips code, numbers and names with their own
casing. Code spans pair backtick runs of equal length and are found once
per pass, which also makes long transcripts linear. A long dash before a
digit is kept as a possible minus sign.
Fillers quoted with corner brackets or full-width parentheses are
mentions and stay. Recapitalization leaves addresses, domains and paths
as written, and a long dash before a decimal point is kept as a sign.
Swap the meter and the processing label in sequence instead of
cross-fading them on top of each other, hold the processing silhouette
through the finish so the meter cannot flash under the delivered check,
and never return to "Transcribing" once the pill has read "Polishing".

Skip becomes a capsule nested in the pill's trailing end. The label,
the chip and the capsule's width now follow one eased value, and the
hover hint fades out with its last words instead of vanishing.
A healthy server finishes cleanup in 170 ms to 1 s, so the label and the
Skip chip flashed past on every Dictation. The pill now names itself
"Polishing" only after 400 ms of cleanup and offers Skip after 1.5 s.
@softmaxe
softmaxe enabled auto-merge (squash) November 12, 2026 06:52
@softmaxe
softmaxe merged commit 29eabd4 into main Nov 12, 2026
7 checks passed
@softmaxe
softmaxe deleted the feat/voice-quality-ports branch November 12, 2026 07:01
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