Focus

import { createCallInput } from 'alkove'

const input = await createCallInput({ voiceprint: localStorage.getItem('print'), learn: true })
pc.addTrack(input.track)
input.on('voiceprint', (p) => localStorage.setItem('print', JSON.stringify(p.toJSON())))

alkove-focus-1-preview isolates one voice in real time, from 16 kHz audio. It keeps the voice closest to the microphone and removes noise, music and other voices. Given a voiceprint from Timbre, it keeps only that person, even when someone else speaks louder or closer.

Size	1.1 MB (int8 weights)
Input	16 kHz mono, 256 samples (16 ms) per step, plus an optional 192-value print
Output	the isolated voice, 16 kHz, one step (16 ms) late
License	Apache-2.0

Two modes, one model

• No print: Focus keeps the closest voice. Nothing to set up. A voice farther from the microphone is treated as interference.
• With a print: Focus keeps the person of the print and removes everyone else, a TV, a colleague or a neighbor, whatever their level.

The input’s mode tells you which one runs: 'focus' or 'focus+print' (or 'raw' while the models load or after a failure).

input.on('mode', ({ mode, reason }) => console.log(mode, reason))

Using Focus from the SDK

Focus is always on in createCallInput() and createAgentInput(). The options that change what it does:

Option	Effect
voiceprint	The print of the person to keep: a Voiceprint, its JSON, or the JSON as a string.
learn: true	Learns the user’s print while they talk (about 6 s of their voice), then keeps only them. Emits learn with the progress and voiceprint when the print changes.
farEnd	What the other side plays (a remote track, your TTS): no learning while it plays.
strength	From 0 (the raw microphone) to 1 (Focus only, the default): a mix of both, delay-aligned.

A new print takes effect at the next silence, so a sentence is never cut in the middle. See Voiceprints for making, storing and learning prints.

For a file, Alkove.process() runs the same model, faster than real time, with the same output as the live path:

import { Alkove } from 'alkove'

const r = await Alkove.process(file, { focus: { voiceprint: print } })
const wav = r.toWav() // the isolated voice
const both = r.toWav('both') // raw on the left, Focus on the right

focus: true keeps the closest voice. Focus always comes with Cue in process(), so r.speech lists the speech segments of the cleaned audio. See Files and streams.

Latency

Focus adds 16 ms (one step) on top of its 16 ms step. A live input reports its total delay in input.latency (milliseconds): Focus, the audio queue and, on a call, the 100 ms look-ahead of the gate that opens when the user starts speaking.

Performance

Native, 1 thread, Galaxy Z Fold8 (Android 17)	about 39 ms per second of audio
Native, 1 thread, Apple M1 Max (macOS)	about 56 ms per second of audio

Between sentences, a live input lets Focus sleep: when its output and the microphone stay quiet, the engine stops running until the microphone rises again. It saves battery and heat on phones.

Benchmarks

Measured with the published file. The model card gives the full tables, with open models run the same way on the same audio.

Test	alkove-focus	Reference
Another voice in the room, Common Voice test sets, 7 languages: SI-SDR of the target, with a print	7.1 dB (no print: 4.5)	raw microphone 3.4 dB
Same scenes: level of the other voice speaking alone, with a print	-22.6 dB (no print: -7.8)	raw 0 dB
Same with DEMAND noise at 10 dB: SI-SDR, with a print	5.0 dB	raw 2.4 dB
AMI meetings, headset microphone (Whisper large-v3 turbo): word error rate, with a print	16.1 % (no print: 16.7 %)	raw headset 20.8 %
AMI: words of other voices that get through, per minute, with a print	0.5	raw headset 17.0
DNS Challenge 3 blind, real recordings: DNSMOS P.808, overall	3.36, 2.70	noisy input: 2.96, 2.33

Limitations

• 16 kHz only, one target voice.
• Made for a microphone close to the speaker (phone, headset, laptop in front of you). With a distant room microphone (several meters), where every voice is far, Focus removes much of the target’s speech too: do not use it there.
• Without a print, the voice kept is the closest one: a farther speaker is treated as interference.
• With a print, when another voice starts on its own, its first ~150 ms can pass, 25 to 35 dB under it.
• Very quiet voices (around -45 dBFS) are kept less well without a print than with one.
• Built for speech; it is not meant to process music.

Out of scope: distant microphones (conference rooms, smart speakers across the room), keeping several speakers at once, music or high-fidelity listening, and isolating someone’s voice in a recording without their consent.

License

Apache-2.0, for the weights. Focus started from the GTCRN checkpoint trained on DNS3 (Rong et al., ICASSP 2024, MIT): its license ships with the model, in NOTICE.

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 256 samples:

Name	Type
audio (in)	float32 [1, 256]	16 kHz mono, -1 to 1
print (in)	float32 [1, 192]	Timbre’s embedding, or zeros to keep the closest voice
audio_out (out)	float32 [1, 256]	the isolated voice, 256 samples late

The engine keeps Focus’s state from one call to the next; clear it when you change the print. The SDK does all of this for you, plus resampling, the audio queue, sleep between sentences and the fallback to the raw microphone.