Cue

import { createAgentInput } from 'alkove'

const input = await createAgentInput()
input.on('speechstart', ({ start }) => agent.userStarted(start))
input.on('speechend', ({ start, end }) => agent.userStopped(end))

alkove-cue-1-preview detects speech: every 30 ms, the probability that someone is speaking, from 16 kHz audio. It is small, runs in real time, and is made to run behind Focus, so the voices and the noise Focus removed do not open it.

Size	1.1 MB (int8 weights)
Input	16 kHz mono, 480 samples (30 ms) per step
Output	probability of speech, one step (30 ms) late
License	Apache-2.0

Using Cue from the SDK

Cue runs behind Focus in every call and agent input. You never call it directly; you get its decisions as events:

Event	When
speechstart	Speech is confirmed. start is when it began, in seconds.
speechend	Speech has stopped. start and end in seconds.

What the recipes do with it:

• A call (createCallInput): the outgoing audio is gated between sentences. The gate opens as soon as speech starts (with 100 ms of look-ahead, so the first syllable is not cut) and closes after it ends.
• A voice agent (createAgentInput): audio stays continuous, nothing is cut; utterance delivers each sentence as one segment, with 300 ms of pre-roll.

Times are measured on the input audio: the delay of Focus and Cue is already taken off.

Tuning the decision

The default decision: speech starts when the probability reaches 0.5 and holds for 0.2 s, and ends once it has stayed under 0.1 for 0.6 s. Change it with the speech option of an input:

const input = await createAgentInput({
  speech: { stopSecs: 0.8 }, // wait longer before ending a turn
})

Field	Default	Meaning
confidence	0.5	probability that opens speech
exitConfidence	0.1	probability under which speech may end
startSecs	0.2	how long speech must last to be confirmed
stopSecs	0.6	how long it must stay under exitConfidence to end

These defaults are tuned behind Focus. Lower stopSecs for snappier turns, raise it if users get cut off while they pause to think.

Files

Alkove.process() and Alkove.stream() give the speech segments of a file:

import { Alkove } from 'alkove'

const { speech } = await Alkove.process(file, { focus: true, cue: { stopSecs: 0.4 } })
// [{ start: 1.52, end: 4.22 }, …]

With focus, Cue runs on the cleaned audio, as in a live input. With cue alone, it runs on the raw audio. See Files and streams.

Cue or Spot

Spot gives the same speech probability as Cue. A wake listener uses Spot’s to know when a word comes after a pause, so it needs no Cue. Cue is for when you need speech detection without keyword spotting, behind Focus.

Performance

Native, 1 thread, Galaxy Z Fold8 (Android 17)	1.7 ms per second of audio
Native, 1 thread, Apple M1 Max (macOS)	2.6 ms per second of audio

Benchmarks

24 test streams of one minute, run behind alkove-focus. In each one, a voice between -45 and -25 dBFS speaks over noise or music, and another voice speaks alone in 40 % of the gaps. Cue uses the default decision above; Silero VAD v6 (MIT) uses 0.65 / 0.5, on the same streams.

	Cue	Silero VAD v6
Missed speech	2.93 %	2.96 %
False starts (24 min)	4	4
Open during the speaker’s silences	6.4 %	5.9 %
Speech onset	10 ms after Silero

Alone behind alkove-focus: no false start on noise for either; 0.1 false start per minute on music (Silero 0.4); open 50 % of the time on TV voices (Silero 81 %).

Speech from Common Voice French dev (CC0) and LibriTTS-R dev-clean (CC BY 4.0); noise and music from MUSAN (CC BY 4.0).

Limitations

• 16 kHz only.
• It reacts to any voice: behind alkove-focus with a print, the other voices are already gone.

Out of scope: telling who is speaking (use Focus with a print first), predicting the end of a turn from meaning (Cue detects speech, not intent), and detecting sounds other than speech.

Running it without the SDK

The model is an .alk file for the Alkove engine (Rust, Apache-2.0), natively or in WebAssembly; its format is documented. One call per 480 samples:

Name	Type
audio (in)	float32 [1, 480]	16 kHz mono, -1 to 1
speech (out)	float32 [1, 1]	probability of speech, one step late

The engine keeps Cue’s state from one call to the next; it starts at zeros.