alkove

On-device voice input for the web: keep the user's voice and drop the rest, detect speech, hear a wake word, recognize a voice. Everything runs in the browser, in WebAssembly. No server, no key, no telemetry.

Preview (0.x). A minor version (0.2, 0.3) may break the API until 1.0; a patch never does. See stability and the changelog.

Read this first. Alkove is made for a near microphone and 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, so on a room microphone in the middle of a meeting table it removes the other participants. Audio is processed at 16 kHz: the output carries speech up to 8 kHz ("HD voice", not full band).

Install

npm install alkove

ESM only, no runtime dependency (the engine, Alkove's own, is bundled). Works without COOP/COEP headers: no SharedArrayBuffer, no threads.

A call (WebRTC)

import { createCallInput } from 'alkove'

const input = await createCallInput({ learn: true, farEnd: remoteTrack })
pc.addTrack(input.track)
input.bind(pc.getSenders()[0]) // the SDK swaps in the raw mic if the audio clock stalls
input.on('voiceprint', (p) => localStorage.setItem('print', JSON.stringify(p.toJSON())))

Out goes the user's voice only, gated between sentences, at a steady level. With learn: true, the user's print is learned while they talk (about 6 s of their voice), then only their voice goes out.

A voice agent (speech to text, LLM)

import { createAgentInput } from 'alkove'

const input = await createAgentInput({ farEnd: 'manual', output: { format: 's16', rate: 24000 } })
input.on('audio', (pcm) => ws.send(pcm)) // continuous, 32 ms packets
input.on('utterance', ({ pcm }) => transcribe(pcm)) // one segment per sentence, 300 ms of pre-roll included
input.farEndActive(true) // the agent starts speaking; false when it stops

Nothing is cut: audio is continuous, speechstart / speechend tell where the user speaks.

• near: true: a voice 10 dB under the user's, a colleague across the room or a video on another device, opens no phrase. No voiceprint: it goes by loudness.
• input.listen(false) mutes (the phrase under way comes as speechcancel), input.hold() / input.release() hold to talk.
• await input.ready once the models run; a quiet microphone is brought up before Focus by itself.

A wake word

import { createWakeListener, Spot } from 'alkove'

const word = await Spot.compile('Camille', { lang: 'fr' }) // or enroll it: Spot.enroll('Camille', { lang: 'fr' })
const wake = await createWakeListener({ words: [word] })
wake.on('wake', ({ word, start, audio }) => …)

Any word, typed in one of 7 languages (fr, en, es, it, pt, ca, ro). By default it counts only after a pause (far fewer false alarms; "Bon, Camille, on y va ?" does not wake it, use context: 'both' for that). Enrolling the word (the user says it 3 times) sets its thresholds on their voice. onlyMyVoice: print ignores other voices.

React

import { useAgentInput } from 'alkove/react'

const { status, error } = useAgentInput({ enabled: on, on: { utterance: ({ pcm }) => transcribe(pcm) } })

useAgentInput, useCallInput and useWakeListener take the options of their function, plus enabled and on. The input opens with the component and closes with it, strict mode and Next.js included. The React guide; two apps in examples (Next.js, Vite).

Files, offline

import { Alkove } from 'alkove'

const r = await Alkove.process(file, { focus: true, cue: true })
r.audio // Float32Array, 16 kHz, processed
r.raw // the same audio, raw
r.speech // [{ start, end }]

The same output as in real time for the same 16 kHz audio. In Node, the same WebAssembly engine as the browser.

Primitives

Timbre (voice prints: embed, verify, identify), Lex (text to phonemes), Spot (compile, enroll), Voiceprint, VoiceprintLearner. Inputs expose mode (raw, focus, focus+print, with the reason when raw), status, latency and typed diag events to plug into your own logs.

Errors

Creating an input fails only when there is no sound at all: AlkoveError with code mic.denied, mic.not-found, mic.busy, page.insecure or browser.unsupported. Anything about models, network or a slow device never fails: the raw microphone goes out at the same delay, mode says why, and the SDK retries.

Models and self-hosting

Models are not in the npm package. By default they are downloaded from Hugging Face (alkove-ai) at a revision pinned by this version, checked by sha256 and cached in the browser (Cache Storage). Good for trying it out. In production, host them yourself: your users then talk to your servers only.

npx alkove models ./public/alkove

Alkove.configure({ modelBaseUrl: '/alkove/', assetBaseUrl: '/alkove/' })

assetBaseUrl also serves the worker and the worklet as files instead of blob URLs. Alkove.preload('call', { onProgress }) downloads and compiles everything ahead (a settings screen), Alkove.cached('call') tells whether it can start offline.

| Usage | Download (brotli) | |---|---| | call / 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 | Alkove's own, 141 KB |

Content Security Policy

• script-src: 'wasm-unsafe-eval', and blob: (the worklet falls under script-src) unless assetBaseUrl is set
• worker-src: blob:, or the origin of assetBaseUrl
• connect-src: your modelBaseUrl, or Hugging Face and its download hosts

Privacy

No network call besides downloading the models, from the origin you choose. No telemetry. Neither audio nor voice prints leave the device. A voice print your app stores or sends is biometric data (GDPR art. 9): treat it as such.

Other platforms

The same models and core, the same outputs to the bit:

• Python, for servers and voice agents: pip install alkove, with plugins for LiveKit Agents (livekit-plugins-alkove) and Pipecat (pipecat-alkove).
• Android, natively in Kotlin: ai.alkove:alkove from maven.alkove.ai, and ai.alkove:alkove-livekit for a LiveKit room.

License

Code and model weights: Apache-2.0. See NOTICE. The Apache license grants no right to the Alkove name or logo: a fork may not call itself Alkove.