Phones and Safari

A phone is the device Alkove is made for: one person, the microphone close to their mouth. Mobile browsers add their own rules about when audio may run. This page lists them and what the SDK does.

Start in a click

Browsers, Safari first, keep an AudioContext suspended until the user touches the page. Create the input from a click or a tap:

startButton.onclick = async () => {
  const input = await createCallInput()
  if (input.status === 'blocked') {
    resumeButton.hidden = false
    resumeButton.onclick = () => input.resume()
  }
  input.on('status', (s) => (resumeButton.hidden = s !== 'blocked'))
}

The SDK opens the microphone first, then creates its context, and tries to start it. If the browser refuses, status is blocked: call input.resume() inside a click handler. On the way back to running, WebKit can pass through interrupted.

If you pass your own audioContext, create it after the microphone is open where you can: Safari can pick a wrong sample rate for a context opened before the capture.

Interruptions

What happens	status	What the SDK does
a phone call or Siri takes the audio (iOS)	interrupted	back to running when the system gives it back
Android suspends the context (audio mode change)	stays running	resumes it by itself; blocked only if that fails
the context cannot restart without a gesture	blocked	waits for input.resume()

The microphone has its own events (mic):

• muted and unmuted: the system took the microphone for a moment. The track comes back by itself.
• lost: the track ended. When the SDK opened the microphone, it reopens it, right away and then after 5 s, 10 s, up to a minute between tries, and emits reopened.
• denied: the permission was withdrawn (from the browser or the system notification). The SDK stops trying.

The output track stays the same through all of this: a WebRTC sender has nothing to redo.

Plugging a headset changes the system’s default microphone; by default the SDK follows it (followDefault). Opening the microphone switches a Bluetooth headset to its call profile; Focus works at 16 kHz anyway.

Hidden pages

Chrome freezes a hidden page after about a minute without sound. Call inputs play a 30 Hz tone at −54 dBFS while running, which phone speakers cannot play, so a call survives in a background tab. Turn it off with keepAliveWhenHidden: false.

Agent inputs and wake listeners play nothing, so they have no such tone.

Battery and heat

The models run in two workers, never on the main thread:

• the real-time worker runs Focus, Cue and Spot on every 16 ms of audio, fed directly by the AudioWorklet;
• the background worker runs Timbre and Lex (learning a print, checking a voice, compiling a word) and Alkove.process, off the audio path.

On call and agent inputs, Focus and Cue go to sleep in silences: after 2 s where Focus’s output is quiet and the microphone sits at its noise floor, they stop computing (diag idle with asleep: true) until the microphone rises 9 dB over that floor. This is always on.

Each model card gives its speed, measured natively on a phone.

A wake listener runs Spot alone on the raw microphone, the whole time it is open.

Slow phones

When a phone cannot keep up, the input does not stutter: it sends the raw microphone at the same delay (mode raw, reason overload) and tries the models again after 10 s, then 20 s, 40 s, up to 160 s. The timing diag, every 2 s, tells you how close to the limit a device runs (meanMs, maxMs, and slow, the steps that took longer than 16 ms). See Errors and diagnostics.

Downloads

The models are fetched once, checked by sha256 and kept in Cache Storage. On a phone, download them ahead on a settings screen, ideally on Wi-Fi:

await Alkove.preload('call', { voiceprint: true, onProgress: ({ loaded, total }) => bar.value = loaded / total })
const offline = await Alkove.cached('call', { voiceprint: true }) // true: it can start without network

Usage	Download (brotli)
call or agent	Focus 0.9 MB + Cue 0.9 MB, + Timbre 0.7 MB with a print
wake word	Spot 2.5 MB, + Lex 1.5 MB for a typed word
engine	134 KB

Serve them from your own origin in production: see Self-hosting.

Browser requirements

• AudioWorklet: Safari 14.1 or later. Without it, creating an input fails with browser.unsupported.
• WebAssembly SIMD: needed by the models. Without it, the input works but stays on the raw microphone for the session (mode raw, reason unsupported, diag engine.unsupported).
• A secure page: HTTPS or localhost, or page.insecure.
• No special headers: no SharedArrayBuffer, no threads, so no COOP or COEP.

Check all of it before you show a microphone button:

const { worklet, simd, secure } = Alkove.supported()

What is tested

The SDK’s browser tests run on the engines of Chromium, Firefox and WebKit (desktop builds, through Playwright): blocked then running after resume() in a click is checked in WebKit, and Firefox’s silent input (no channel at all while nothing plays) is handled as silence. The phone figures above come from Chrome on Android, where the models give the same outputs, to the bit, as on a desktop. Try your app on the phones your users have, Safari on iOS included, before you ship.