“Can’t I just use ElevenLabs?”
ElevenLabs offers a broad voice platform. Converse focuses on the complete live conversation: turn-taking, interruption handling, tools and predictable usage pricing.
Converse
The realtime voice layer
Talking to
machines
still sucks.
Converse is the realtime voice API that finally gets conversations right.
Speech recognition, turn-taking, tools and speech output — running together in one low-latency session.
No card. Start talking.
feels natural.
A natural spoken exchange in which Converse responds to the user, yields immediately when interrupted, and continues with the corrected request.
Live conversation
Speak whenever you like
You
Can you move my booking to Friday?
Converse
Of course — I’ll move your booking to Fri—
Interrupts naturally
Actually, make that Monday.
Converse
Monday it is. Anything else?
See test dates, model versions and complete results in the benchmark tables.
Why Converse
Built for natural
conversation.
The messy parts of spoken interaction belong in the platform, not in application code.
-
Real turn-taking
Barge-in, backchannels and pauses handled the way people actually talk.
-
One persistent session
Speech recognition, the language model, tools and speech output orchestrated together.
-
Low latency
Engineered for flow. No awkward gaps, no press-1-to-continue.
-
Distinctive voices
Warm, recognisable voices designed for clear and consistent customer experiences.
Start talking. Interrupt.
Take your time.
Open the playground →
Browser SDK
One clean API.
Own the instructions and tools. Converse owns the latency-critical conversation loop.
import { ConverseClient } from
'@trelis/converse';
const client = new ConverseClient({
url: 'wss://converse.trelis.com/ws',
sessionId: credential.session_id,
apiKey: credential.api_key,
mode: { kind: 'converse' },
});
await client.connect();
await client.startMic();Measured in public
Hold us to it.
We test each voice system with the same spoken prompts and publish the dates, model versions and complete results. We measure interruption handling alongside speed because a fast reply that cuts someone off is still a poor conversation.
Explore the benchmarksVoice-to-voice response latency p50 · 2026-07-07
Converse837 ms
Gemini Live 3.11,269 ms
OpenAI Realtime1,846 ms
Gemini Live 2.52,962 ms
Response latency runs from the end of user speech to the first audible assistant audio.
Simple pricing
$3 per conversation hour
Connected time is billed per second. New accounts start with $3.00 of credit — about an hour to explore, with no card required.
- Partial seconds round up
- One concurrent session by default
- Two-hour maximum per connection
Your data
Clear terms, plain language.
Converse is in alpha. Conversations are recorded and may be used for quality assurance, evaluation and model training.
Access
Invite-only, for now.
We’re scaling capacity deliberately. Building support or device voice that needs real turn-taking? Tell us about it.
“Can’t I just use OpenAI Realtime?”
It’s a strong model API. Converse is built around the voice experience itself: measured turn-taking, tools, distinctive voices and predictable pricing.