How it works

Everything below your agent’s logic, on your machine.

Calls arrive over SIP or WebRTC. Synauson mixes and routes the audio, hears when someone speaks and when they have finished, writes down what they said and speaks your agent’s reply. Your agent sees one ordered stream of events and sends commands back. It never has to touch audio.

Calls enter your servers over SIP or WebRTC. Synauson handles the media and speech AI and exchanges events and commands with your agent. Only text goes to the language model you choose. Phone calls SIP Browsers, apps WebRTC Your servers synauson Media engine Voice activity and turns Speech-to-text Text-to-speech Optimized models Your agent your code events commands Your choice of LLM text Calls enter your servers over SIP or WebRTC. Synauson handles the media and speech AI and exchanges events and commands with your agent. Only text goes to the language model you choose. Phone calls (SIP), browsers and apps (WebRTC) Your servers synauson Media engine Voice activity and turns Speech-to-text Text-to-speech Optimized models events commands Your agent your code text Your choice of LLM
Call audio stays inside the dashed line. Your agent sends text to the model you choose, and Synauson sends us only license and usage data.

What’s inside

One engine for the media and the speech AI.

Media engine

SIP and WebRTC calls, multi-party conferences and recording, with no cap on conferences or participants.

Voice activity and turn detection

Hears when someone starts speaking, and tells a pause from the end of their turn, so your agent replies at the right moment.

Streaming speech-to-text

Transcribes English as the caller talks. Words arrive moments after they’re spoken, with timings.

Text-to-speech

Your agent sends text and Synauson speaks it into the call in one of 20 American English voices, with no audio pipeline for you to build.

Runs on the hardware you have

Every model runs on standard x86_64 CPUs, from a laptop to a server. Speech-to-text workers on a machine share one copy of the int8 weights, and at startup Synauson measures the machine and takes on only the streams it can transcribe in real time.

Two ways to run it

Embed the engine in your own process with an in-process SDK. Today that’s jsyn for Java on Windows and Linux (laptops included), and more languages are planned. Or run the gRPC server in a container or under systemd on Linux; its public API comes with API 1.0 in the next release. Either way, every instance shares one license.

Why local

The future of AI is local.

Models keep getting smaller and better, and inference keeps getting cheaper. The companies that run their own will own the experience and the cost curve.

Calls stay private

Audio is processed on your machines and never streamed to an AI vendor. Security reviews get shorter, and data stays where your customers expect it.

Conversations feel natural

Hearing, turn-taking and speech all happen on your machine, so no turn waits on a speech service in a distant data center.

Costs you can predict

Synauson is licensed by concurrency: you pay for how many AI sessions run at once, not for every minute of every call.

Yours to own

Run it on your own servers or in your own cloud account, under your own name, with no Synauson branding required. Shipping it to customers to install needs an OEM agreement.

Free plan

Start free, and stay free.

2 AI sessions at once, every AI feature, and you can use it in production. No credit card required.

Get a free keyCompare plans