Voiceprint recognition in 2 seconds: what short-utterance verification changes

2 seconds of speech is all our voiceprint model needs to recognize your unique voice.

To see what that means in practice, let's get the ruler out:

"Hey, how have you been?" — 1.6 s

"When is my meeting with Peter this afternoon?" — 2.8 s

Notice the second one is longer than the model needs. Verification would already be done before you finish asking.

That matters because longer utterances do improve accuracy — more audio, more evidence. That's why enrollment still takes 6 seconds: it's a one-time setup, and worth getting it right.

But every recognition after that has to fit inside a normal sentence. If each authentication demanded 6 seconds, you'd be reciting a paragraph to unlock something — extra hassle and awkward in a lot of situations.

At 2 seconds, verification runs quietly in the background and completes before most sentences do.

MLSP voiceprint model makes confident decision from just 2 seconds of audio. Here's what that looks like — genuine user and impostor:

In the positive match example on the left, different sentences are used for enrollment and test, in other words, it’s sentence independent. The negative match example on the right, showcases how the model output, in probability from 0 to 1, changes as more audio is collected.

The interesting question isn't whether 2 seconds is fast enough — it's what becomes possible once verification is effectively free. If you're designing around that assumption for your product, I'd like to compare notes.