Spot

import { createWakeListener, Spot } from 'alkove'

const word = await Spot.compile('Nova', { lang: 'en' })
const wake = await createWakeListener({ words: [word] })
wake.on('wake', ({ word, score }) => console.log('heard', word, score))

alkove-spot-1-preview detects any word you type, in real time, from 16 kHz audio: a wake word, a command, a product or contact name, a word in the middle of a sentence. No model to train, no recordings to collect. It also says when someone is speaking.

Size	2.8 MB (4-bit and int8 weights)
Languages	French, English, Spanish, Italian, Portuguese, Catalan, Romanian
Input	16 kHz mono, 480 samples (30 ms) per step
Output	phoneme log-probabilities, 90 ms late; probability of speech, 30 ms late
License	Apache-2.0

How a word is found

Spot hears phonemes, not words. To listen for a word, the SDK:

1. turns the text into its likely pronunciations with Lex (Spot.compile);
2. searches those phoneme sequences in what Spot hears, continuously;
3. fires when the score of a pronunciation crosses the word’s threshold.

const word = await Spot.compile('Camille', { lang: 'fr' })
word.toJSON() // text, languages, pronunciations with their probabilities, thresholds

A compiled word is plain JSON: store it, ship it with your app, compile it on a server-side build step if you like. Pass the word, its JSON or the JSON string to any API that takes words. You can also skip Lex and write the phonemes yourself, between slashes.

lang takes one language or several (['fr', 'en']) for a name said in more than one language.

Thresholds

Every word has two thresholds, one for anywhere in speech and a lower one for after a pause. Where they come from:

• Compiled (Spot.compile): a default per language, set from typed words on the published Spot. Good enough to try any word right away.
• Enrolled (Spot.enroll): the user says the word a few times (3 by default), and the thresholds are set from these takes and from what is heard between them. The word is then tuned to their voice and their microphone.

const enrollment = Spot.enroll('Camille', { lang: 'fr' })
enrollment.on('take', (take) => showTake(take))
const { word, voiceprint } = await enrollment.done

Enrolling also gives a voiceprint of the user, ready for onlyMyVoice. See Wake words.

Where a word counts

The context option decides where in speech a word may fire:

context	Fires on
'pause'	the word said after a pause (300 ms without speech). Far fewer false alarms: “so, Camille, let’s go” does not wake it.
'anywhere'	the word anywhere, including mid-sentence.
'both'	either, each with its own threshold.

Each wake event says which one fired (context: 'pause' or 'anywhere'), with its score, the threshold it crossed, and start and end in seconds.

Using Spot from the SDK

API	Spot reads	Default context
createWakeListener({ words })	the raw microphone	'pause'
input.spot(words) on a call or agent input	Focus’s output, in the same worker	'anywhere'
Alkove.process(file, { spot: { words } })	Focus’s output with focus, the raw audio without	'anywhere'

A wake listener runs Spot alone, the lightest way to listen all day. With onlyMyVoice: print, a detection counts only if Timbre recognizes the user’s voice in the word.

On a call or agent input, input.spot() reads the cleaned audio: no second microphone, no second Focus, and with a print the other voices are already gone. A word enrolled on the raw microphone keeps thresholds measured there; to listen behind Focus, enroll it on the input with input.enrollWord().

Performance

Native, 1 thread, Galaxy Z Fold8 (Android 17)	about 7.5 ms per second of audio
Native, 1 thread, Apple M1 Max (macOS)	about 13 ms per second of audio

Benchmarks

Recall of typed words at 1 and 0.5 false alarm per hour. 40 words per language, each phonetized by alkove-lex alone (its first pronunciation, plus the variants it gives at least 25 %). For each word, the threshold is set for the target false-alarm rate on half of the speakers, and recall is measured on the other half. False alarms are counted on all the speech that does not contain the word.

Recall	fr	en	es	it	ca	pt	mean
at 1 false alarm / hour	89.9	66.9	69.9	78.8	91.3	85.3	80.4 %
at 0.5 false alarm / hour	85.3	62.2	61.9	74.0	86.8	72.2	73.7 %

Only after a pause (the word said after 300 ms without speech), at 1 false alarm per hour: 85.9 %, at 0.5: 81.3 % (mean of fr, en, es, it, ca, pt).

French, English, Spanish and Italian on MOCKS (CC BY 4.0 / MPL 2.0); Catalan and Portuguese on the Common Voice test sets (CC0). Romanian is supported, but its Common Voice test set is too short to set a threshold at these false-alarm rates, so it is left out of the table.

Limitations

• 16 kHz only; the 7 languages above.
• As with any keyword spotter, short words false-alarm more than long ones.
• Recall is lower in English than in the other languages.

Out of scope: transcription (Spot finds the words it is given, it does not write down what was said), languages outside the 7 above, and triggering anything safety-critical or irreversible on a single detection: false alarms happen, at the rate you choose.

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
logp (out)	float32 [1, 1, 59]	log-probabilities of the CTC blank and the 58 phonemes, 3 steps (90 ms) late
speech (out)	float32 [1, 1]	probability of speech, 1 step (30 ms) late

The engine keeps Spot’s state from one call to the next; it starts at zeros. The model gives phoneme probabilities; the search for a word and its thresholds are in the SDK.