Timbre

import { Timbre } from 'alkove'

const timbre = await Timbre.create()
const print = await timbre.embed(enrollmentPcm) // 16 kHz Float32Array, a few seconds of the user
const heard = await timbre.embed(lastSentencePcm)
timbre.verify(print, heard) // { same, score }

alkove-timbre-1-preview turns speech into a voiceprint. It answers “is this my user?” on the device, and gives Focus the print of the person to keep.

Size	0.74 MB (int8 weights)
Input	16 kHz mono speech, one window of 280 to 24,000 samples (1.5 s, advised)
Output	two unit vectors of 192 values: embedding (the print Focus follows) and id (to compare voices)
License	Apache-2.0

Using Timbre from the SDK

Most apps get voiceprints without calling Timbre:

Where	What Timbre does
learn: true on a call or agent input	learns the user’s print while they talk (about 6 s of their voice), then Focus keeps only them; voiceprint events give you the print to store
onlyMyVoice: print on a wake listener	checks that the voice that said the word is the user’s (adds 0.5 s of latency to a detection)
Spot.enroll()	gives a print from the takes of the word, along with the word
VoiceprintLearner	an enrollment screen (“read this sentence”), with the same rules as learn

Call it directly to compare voices yourself:

API	What it gives
Timbre.create()	a Timbre instance, once the model is loaded in the background worker
timbre.embed(pcm, { source })	a Voiceprint from 16 kHz audio. source is 'raw' (the default, a microphone or a file) or 'focus' (Focus’s output).
timbre.verify(print, heard, { level })	{ same, score }: whether heard is the voice of print
timbre.identify(voice, known, { level })	{ who, score } for the best match among named prints, or null

embed() cuts the audio into 1.5 s windows and averages them, leaving out the 10 % least alike, so a cough or a word from someone else does not spoil the print. Give it at least 1.5 s; a few seconds of the user alone make a good print.

Levels

verify() and identify() take a level:

level	Use it when
'lenient'	rejecting the user would be worse than letting another voice through
'balanced' (default)	as many wrong rejections as wrong acceptances
'strict'	letting another voice through would be worse than asking the user to repeat

The threshold behind each level adapts to how much voice the print was made from and how much was heard, so a short word and a long sentence are both judged fairly. The first argument of verify() is the reference print, the second what you heard.

Storing prints

import { Voiceprint } from 'alkove'

localStorage.setItem('print', JSON.stringify(print.toJSON()))

const saved = Voiceprint.fromJSON(localStorage.getItem('print')!)
if (!saved) relearn() // made by another version of Timbre

A print is plain JSON: numbers, no audio. Voiceprint.fromJSON() returns null for a print made by another version of Timbre, which no longer means anything to this one: learn it again. Every API that takes a print accepts the Voiceprint, its JSON, or the JSON string.

Voiceprint.merge([a, b]) combines prints of the same person (a phone and a headset, say). A print is personal data: see Voiceprints and Privacy for consent and storage.

Performance

Native, 1 thread, Galaxy Z Fold8 (Android 17)	2.0 ms per 1.5 s window
Native, 1 thread, Apple M1 Max (macOS)	2.9 ms per 1.5 s window

Timbre runs in the background worker, never on the audio path.

Benchmarks

Equal error rate (same voice or not, lower is better) on the Common Voice test sets of 7 languages (CC0): 2,029 speakers never heard in training, 4 clips each (8,057 clips). For reference, the open 3D-Speaker CAM++ (Apache-2.0, 28 MB) on the same trials.

EER	alkove-timbre id (0.74 MB)	CAM++ (28 MB)
A print of about 12 s (3 clips) against a 1.5 s window	2.95 %	1.72 %
The same, impostors only with a close pitch (±2 semitones)	4.11 %	2.32 %
Clip against clip, whole clips (about 4 s)	4.42 %	2.61 %
Clip against clip, 3 s	4.90 %	3.02 %
Clip against clip, 1.5 s	7.14 %	4.75 %

Mean of the 7 languages. Print against a 1.5 s window, per language: French 2.68 %, English 3.20, Spanish 2.79, Italian 2.57, Portuguese 4.28, Catalan 2.83, Romanian 2.30 (30 speakers only). The published file takes windows of up to 1.5 s; the whole-clip and 3 s rows ran the same weights on longer windows.

Harder, and closer to a live call: 1,424 windows taken from scenes with noise, music or another voice.

EER	raw mic	after Focus, with print	after Focus, without print
id, 1.5 s windows	11.9 %	10.1 %	11.6 %
id, 1 s windows	14.4 %	14.2 %	13.4 %
embedding, 1.5 s windows	15.7 %	18.6 %	19.1 %

About ±1 point on this smaller protocol. Speech from Common Voice French dev (CC0), Multilingual LibriSpeech French dev (CC BY 4.0) and LibriTTS-R dev-clean (CC BY 4.0). Noise and music from MUSAN (CC BY 4.0), the DNS Challenge (CC BY 4.0) and RIRS_NOISES (Apache-2.0).

Limitations

• 16 kHz only; short windows (under 1 s) are less reliable.
• embedding is made for Focus; use id to compare voices. The SDK does this for you: verify() and identify() compare id.

Out of scope:

• authentication or access control: at about 3 % equal error rate it is a convenience filter, not a security check;
• identifying or tracking people, or forensic speaker comparison;
• comparing voices across very short windows (under 1 s).

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. Give the engine the window’s length first:

Name	Type
audio (in)	float32 [1, samples]	16 kHz mono, -1 to 1, 280 to 24,000 samples; 1.5 s (24,000 samples) advised
embedding (out)	float32 [1, 192]	unit norm: the print for Focus
id (out)	float32 [1, 192]	unit norm: compare two voices by the dot product

To enroll someone, average the embedding (or id) of several windows of their voice and normalize the result again.