[ANT1OFF]

What

Letting go of dictation no longer clips the last word.

• On a stop (release or toggle), capture runs on for 200 ms, Hex's POST_RELEASE_TAIL. The status changes to "Finishing…" at once.
• The worklet then flushes its partly filled 100 ms frame, which used to be lost. The page sends stop only after the worklet confirms, with a 100 ms guard in case the audio thread has stopped answering.
• Frames keep flowing during stopping, and the level bars stay at 0.
• Escape, and a stop that overtakes the start (a tap that never got to listen), get no tail.

The speech process already accepts frames of 1 to 16000 samples, so nothing changes on its side.

Results

Real recogniser, synthetic voice. I ran 8 sentences from macOS say through the installed Nemotron int8 560 ms model and the real createSpeechService, at 3 frame phases, so 24 runs per row. "Before" is full 100 ms frames up to the release. "After" is the release + 200 ms plus the short frame. Audio was framed by a scratch harness, not captured through the app; the script is not committed.

Last word kept:

Release vs end of speech	Before	After
−200 ms	0/24	24/24
−100 ms	10/24	24/24
−50 ms	16/24	24/24
0 ms	22/24	24/24
+100 ms	24/24	24/24
+200 ms	24/24	24/24

Whole sentence exact: 22/24 at every offset after, 0–22/24 before. The two misses after are one sentence at two of its three phases: "Run the" at the start heard as "runda" or "rhonda". The same error appears at +200 ms before this change, so it is a recogniser error at the first words and not clipping at the end.

Tests

• capture.test.ts replays a synthetic WAV (tone "words") through the real worklet class, startCapture and session over fake Web Audio, released at −100 to +200 ms. Every sample up to the end of speech arrives.
• Mutation check: with the tail set to 0, the −100 and −50 ms cases fail. With the flush removed, the short-frame case fails.
• session.test.ts covers Escape during the tail, a stop during the start, a double stop, an error during the tail, and a cancel while the flush is out.

Real worklet in Chromium (fake mic device, mock speech process). A one-off spec, not committed, logged the last frames: 1600, 1536, stop, then 704, 960 and 256 on later runs, so the flush works on a real audio thread. Stop click to idle took 258–337 ms.

Commands: vitest run --project web apps/web/src/dictation (31 passed), e2e dictation.spec.ts (12 passed), typecheck, fmt:check, check:scaffold, oxlint src/dictation.

Known limits

• Not tested with a real microphone or a human voice. The figures above are from say audio.
• Slower finish: text lands about 200 ms later than before, which is the cost of the tail.
• Send during the tail: the thread composer doesn't stop dictation on send (the drafts do), so pressing Enter straight after letting go can send before the last words commit, and they land in the emptied draft. That already happened; the window is now 200 ms longer.
• Press during the tail: a new press while stopping does nothing, as before, and that window is also 200 ms longer.

No visible UI change, so no screenshots.