Sro hears what a conversation means while it is still happening.

Ask any question about live or recorded audio. Sro answers with a probability, twice a second.

Sro is an audio decision model. You write a question in plain English and choose the answer you want back: yes or no, one of several options, or a score. Sro answers from the voice, the timing and the room, with no transcript in between.

{"type": "decision",
 "at": 10.0, "span": 7.0,
 "values": {
   "asked":  0.981,
   "over":   0.032,
   "phone":  0.946,
   "tone":   "warm"
 }}

Ten seconds into a support call, as the agent asks what went wrong.

Hear the Call

160msper decision
on one GPU
90%spoken intent,
zero examples
2×decisions
every second

Live audio, in real time

Stream a call over a WebSocket. Every half second Sro answers each of your questions about the last seven seconds, starting one second into the call.

Recorded calls and clips

Upload any common audio or video file and ask up to 24 questions in a single call. The recording stays available for a day, so you can come back with new questions.

Questions nobody trained it on

Sro never learned the question "Is there a slam?" and still answers it correctly 91% of the time. Any question without a published measurement comes back labelled, so you can test it on your own audio.

manifesto

to everyone building something that listens

voice agents learned to talk before they learned to listen.

most of them still hear a transcript: the words, with everything that made them mean something stripped away.

a transcript misses the pause before an answer, the laugh that turns a complaint into a joke, and the caller who says "i'm fine" in a voice that is clearly not fine.

it keeps the words and throws away the person.

people never listened that way. we hear the tone before the words, and we start deciding before the sentence ends.

sro is built to hear the same way.

you ask it a question in plain english, and it answers from the sound itself with a probability you can act on, twice a second, while the conversation is still going.

ask whether the caller is upset, whether they just asked for something, or whether two people are talking over each other, and the answer arrives while they are still speaking.

your agent gets a number it can act on, with nothing in between to parse.

we started building sro in september 2026. it has been answering live audio since the first week of october.

every number we publish was measured on audio sro never trained on, and any question without a measurement comes back marked as one.

the machines people talk to should hear them properly. that is the whole project.

we call it sro.
from the sanskrit śru, to hear.

api keys are open now. get yours here
- the sro team