Overview

Alkove turns a microphone into clean voice input, on the device. The alkove npm library runs five small models in WebAssembly, in the browser or in Node; the alkove Python package runs them natively, with plugins for LiveKit Agents and Pipecat, and so does the Android library, in Kotlin. No server, no key, no account, no telemetry: audio and voiceprints never leave the device.

import { createCallInput } from 'alkove'

const input = await createCallInput()
pc.addTrack(input.track) // only the user's voice goes out

Read this first

A near microphone, one person

Alkove is made for a microphone close to one person: a phone, a headset, a laptop in front of you. It keeps the closest voice (or the voice of a print) and removes the others. On a room microphone in the middle of a meeting table, it removes the other participants too.

• 16 kHz voice. The models work at 16 kHz. The output track runs at the rate of your AudioContext (usually 48 kHz), but it carries speech up to 8 kHz: HD voice, not full band.
• Latency. A call input adds about 150 ms, of which 100 ms is the gate opening ahead of each sentence. An agent input, without a gate, adds about 50 ms. input.latency gives the real figure, in milliseconds, updated as it moves.
• It never leaves you without sound. When the models are still loading, fail to load or the device is too slow, the raw microphone goes out at the same delay and the SDK tries again. input.mode tells you which path is live and why.

The five models

Model	What it does	Used by
Focus	Keeps the closest voice, or one person’s voice given their print, and removes noise, music and other voices	call and agent inputs, Alkove.process
Cue	Tells when someone speaks, every 30 ms	call and agent inputs (speech events, gate, utterances)
Spot	Hears any word you type, in 7 languages, with no training	wake listener, input.spot()
Lex	Turns a written word into its likely pronunciations	Spot.compile for typed words
Timbre	Makes voiceprints and compares them	learn, onlyMyVoice, Timbre

The models are not in the npm package. They are downloaded once, checked by sha256 and cached; you can host them yourself. See the model pages for what each one can and cannot do.

Pick a recipe

Each use has one function. Start from the one that matches your app.

You are building	Start with	Quick start
A call between people (WebRTC)	createCallInput	A call
A voice agent (speech to text, LLM)	createAgentInput	A voice agent
A wake word, or commands	createWakeListener and Spot	A wake word
Processing recordings	Alkove.process, Alkove.stream	Files
Your own pipeline	Timbre, Lex, Spot, Voiceprint	Primitives
A voice agent on a server	pip install alkove, or a plugin	Python, LiveKit Agents, Pipecat
An Android app	ai.alkove:alkove, AgentInput, WakeListener	Android

A call or agent input can also listen for words behind Focus with input.spot(): one microphone, one Focus, no second input.

How it runs

• An AudioWorklet resamples the microphone to 16 kHz and keeps a short queue. It never runs a model.
• A real-time worker runs Focus, Cue and Spot frame by frame. A background worker runs Timbre and Lex. Both share one compiled engine, and there is one engine per page whatever the number of inputs.
• The audio path never goes through your main thread. Only what you ask for comes back to it: events, and the PCM of an agent input.
• The engine is Alkove’s own (134 KB), bundled in the package: the same result to the bit in every browser. It needs WebAssembly SIMD but no special headers: no COOP / COEP, no SharedArrayBuffer.

What an input tells you

Every input has the same events, with on(type, fn) returning its unsubscribe function:

input.on('mode', ({ mode, reason }) => console.log(mode, reason)) // 'focus', or 'raw' with why
input.on('speechstart', ({ start }) => console.log('speech at', start))
input.on('diag', (d) => myLogger[d.level](d.code, d))

• mode: raw, focus, focus+print, spot or spot+print. When raw, reason is off, loading, failed, overload or unsupported.
• status: starting, running, blocked (the browser wants a user gesture: call input.resume() in a click), interrupted or disposed.
• diag: typed diagnostics (code, level, message) to send to your own logs.

Creating an input fails only when there is no sound at all: no permission, no microphone, an insecure page, an unsupported browser. Everything else is a mode and a diag. See Errors and diagnostics.

Next steps

1. Install the package and check the browser.
2. Follow the quick start for your use: call, voice agent, wake word or files.
3. Before production, self-host the models and set your Content Security Policy.
4. Read Privacy if you store voiceprints: they are biometric data.