Models

Alkove is five small models, each doing a single job, and one engine that runs them on the device: in WebAssembly in the browser, natively in Python and on Android. You rarely load a model by hand: the recipes (createCallInput, createAgentInput, createWakeListener) and Alkove.process() load what they need.

The five models

Model	Job	Size	Languages
alkove-focus-1-preview	Voice isolation: keeps the closest voice, or the voice of a print, and removes noise, music and other voices	1.1 MB	any (speech)
alkove-cue-1-preview	Voice activity: the probability that someone is speaking, every 30 ms	1.1 MB	any (speech)
alkove-spot-1-preview	Keyword spotting: any typed word, no training	2.8 MB	fr, en, es, it, pt, ca, ro
alkove-lex-1-preview	Pronunciation: a written word to its likely phonemes, with probabilities	1.8 MB	fr, en, es, it, pt, ca, ro
alkove-timbre-1-preview	Voiceprints: who is speaking, and the print Focus follows	0.74 MB	any (speech)

All five together weigh under 8 MB. All take 16 kHz mono audio (Lex takes text). The SDK resamples the microphone or your file for you.

The engine, alkove-engine.wasm (466 KB, 134 KB brotli), is Alkove’s own: it runs these five models, with the core that turns their outputs into features, and gives the same result to the bit on every target. It ships inside the npm package, runs on one thread, and needs WebAssembly SIMD. The engine and the .alk format are open: the engine, the format.

Who runs what

You use	Models loaded
createCallInput(), createAgentInput()	Focus and Cue; Timbre too with learn: true (a voiceprint you pass in needs no Timbre)
input.spot(words) on a call or agent input	Spot, reading Focus’s output
createWakeListener()	Spot on the raw microphone; Timbre too with onlyMyVoice
Spot.compile('word')	Lex, to turn the text into phonemes (no model for phonemes written between slashes)
Spot.enroll()	Lex (for typed text), Spot, and Timbre (the takes also give a voiceprint)
Alkove.process()	focus → Focus and Cue, cue → Cue, spot → Spot
Timbre.create(), VoiceprintLearner.create()	Timbre
Lex.create()	Lex

In a live input, Focus, Cue and Spot run in a real-time worker, fed by an AudioWorklet. Timbre, Lex and Alkove.process() run in a second, background worker, so a voiceprint, a pronunciation or a file never delays the live audio. One engine serves the whole page, whatever the number of inputs.

To download models ahead (an onboarding screen, a settings page) and to check that a usage works offline:

import { Alkove } from 'alkove'

await Alkove.preload('call', { voiceprint: true })
const offline = await Alkove.cached('wake', { typedWords: true })

See Self-hosting models for where the files come from and how they are cached.

Speed

Measured natively on a Galaxy Z Fold8, one thread:

Model	Cost
Focus	about 39 ms per second of audio
Cue	1.7 ms per second of audio
Spot	about 7.5 ms per second of audio
Timbre	2.0 ms per 1.5 s window
Lex	about 2.9 ms per word

On that phone the outputs are the same, to the bit, as on a desktop. Timbre and Lex only run now and then (a voiceprint update, a word compiled), never per audio frame.

Versions

Each version of the SDK pins its set of models by SHA-256. A file whose hash does not match is refused (models.hash), so the SDK always runs exactly the models it was tested with, wherever you host them.

When a model changes, the SDK version changes with it, and the changelog says what to redo:

• Timbre changes: voiceprints made by the previous Timbre no longer mean anything. Voiceprint.fromJSON() returns null for them: learn the print again (about 6 s of the user’s voice with learn: true).
• Spot changes: compiled words carry the Spot they were compiled for. A word from another Spot (another name or other weights) is refused with an AlkoveError word.outdated: compile its text again, or enroll it again.

The -preview suffix means the weights and their interface may still change before version 1. See stability and versions.

License

Apache-2.0, for the weights and the SDK. The model cards on Hugging Face (huggingface.co/alkove-ai/<model>) give the full benchmarks and their datasets.

Limits shared by all models

• A near microphone, one person. Alkove is made for a phone, a headset or a laptop in front of you. On a room microphone in the middle of a meeting table, Focus keeps one voice and removes the other participants.
• 16 kHz. The models work on 16 kHz audio: the output carries speech up to 8 kHz (“HD voice”), not full band.
• Speech. The models are built for speech, not music.