[ANT1OFF]

What

Dictation starts sooner, and with an opt-in setting it hears the moment before you press.

• (a) Overlap. getUserMedia now starts at the same time as the session opens. Frames heard before the port arrives are queued and sent after the lead silence, so a slow open (a cold model, IPC) costs no audio. The installed-model lookup is remembered, but only for the slider's own model and only until Settings changes what's installed or the speech process says model-missing, so a fallback never sticks. The microphone permission is remembered once granted; getUserMedia still reports a later revoke. A press therefore no longer waits on up to three speechStatus IPCs and a permission IPC before it opens the mic.
• (b) Graph kept. One AudioContext and worklet per window, built at the composer's focus (beside the model prewarm) and kept suspended between dictations, so no audio device stays open. The worklet gathers only between start and stop, so no part-frame carries over to the next dictation.
• (c) Keep microphone ready, opt-in and off by default. After a composer's focus or the last dictation, it holds the stream for 2 minutes, then lets go. Each focus or dictation restarts the 2 minutes, a dictation in progress never times out, and turning it off, or turning voice input off, lets go at once. While held, it keeps the last 4 frames plus the worklet's part-frame (400–500 ms) in memory only, replaced as they age, and hands them to a press as pre-roll. Focus never makes the first permission ask; that stays with a press. A stream that ended while held (device unplugged) is reopened on the press.
• Settings copy. The row's always-visible detail says the mic stays open for 2 minutes and the macOS mic indicator stays lit ("your system's" off macOS). With it on, the Microphone row no longer says "listens only while you dictate".
• Usage description changed. NSMicrophoneUsageDescription said "only while you dictate", which is untrue with the opt-in on. It now reads: "Gram uses the microphone while you dictate into the composer, and between dictations only if you turn on Keep microphone ready. Speech is transcribed on this Mac and never leaves it."
• (d) Model idle and prewarm: no change, measured. The model loaded in 3.6–4.0 s here (load average ~30; service.ts says about 1 s on an idle machine) and the process sits at ~0.8–0.95 GB RSS loaded. With (a), a cold model no longer loses audio: the speech process queues frames until ready, so a cold start only delays the text. Keeping ~0.9 GB resident for longer than the current 5-minute SPEECH_IDLE_MS, or loading it on every window focus, isn't worth that. Composer-focus prewarm stays as it is.

Latency: press to first frame

Dev app (Vite + mock speech process) in Playwright Chromium. A press of ⌘⇧Space is timed to the first non-silence frame reaching the session port. 12 presses per run, the first (cold) press dropped. The real-mic rows are 2 rounds alternating before and after; the fake-device rows are one run each, not alternated. The machine was heavily loaded: load average 30–60 from other checkouts.

	min	p25	p50	p75	max
Before (real mic)	409	458	496	634	1339 ms
After, default (real mic)	319	351	540	773	1401 ms
After, Keep microphone ready (real mic)	27	32	36	43	68 ms
Before (Chromium fake device)	200	323	366	434	490 ms
After, default (fake device)	148	156	162	271	345 ms

• Default path. The reliable change is that audio is now captured from the moment the mic opens, however long the session takes. Press to first frame on the real mic is not measurably faster under this load. A standalone Chromium microbenchmark explains it: getUserMedia alone takes ~270–290 ms (median) on this Mac's built-in mic, and the audio graph after it only ~60–80 ms, whether built per press or kept. What (a) and (b) remove (the IPCs and the graph build) is small next to that, and the run-to-run noise at load 30–60 hides it. I expect a gain of tens of ms on an idle machine, but I haven't shown it.
• Keep microphone ready. The first frame arrives in ~36 ms and carries the ~0.4–0.5 s before the press, so a word that starts with the press is not clipped.

Verification

• Real microphone: the MacBook Pro mic through Playwright Chromium (--use-fake-ui-for-media-stream only grants the permission) for the latency rows above. Recognition used the mock process, not the model, and nobody spoke. I did not check the macOS mic indicator by eye or run the packaged app.
• Unit: capture.test.ts runs the real worklet over fake Web Audio:
  • the graph is warmed once, stays suspended and no mic is opened on focus
  • no carried-over part-frame
  • pre-roll order and size, and the 2-minute hold
  • a focus extends the hold, turning it off releases at once
  • no timeout under a dictation, an ended track reopens, and a takeover between composers
• Unit: session.test.ts covers frames queued while the session opens, a cancel during the open releasing the mic at once, and each side released when the other fails. use-dictation.test.ts covers the remembered target (never a fallback or a no, forgotten on a lifecycle change or model-missing). settings.test.ts covers the new setting.
• e2e dictation.spec.ts, 14 passed:
  • new: the mic is not opened on focus, opens on the press and closes on the stop
  • new: Keep microphone ready holds the mic from focus, gives the press ≥ 4 pre-roll frames, keeps the stream for the next press, and turning it off in Settings stops the track
• Commands: vitest dictation + settings + contracts, typecheck, lint, fmt:check, check:scaffold, check:docs, e2e dictation.spec.ts.

Known limits

• Default-path latency is not shown to improve beyond noise (see above).
• With the opt-in on, a composer focus opens the mic (indicator lit) even if you don't dictate. That is the trade the setting describes.
• The hold timer is per window; two windows each hold their own stream.

Screenshots

Settings → Advanced → Voice input, with voice input on (dark)

Before	After