The duplex speech model for revenue calls.
MetaVoice listens while it speaks, so calls stay natural through interruptions, overlap and background speech.
Fewer people hang up. More convert.
problem
30-75% of people hang up on voice agents within 30 seconds.
Every hang-up is a lost booking, payment or customer. Some become one-star reviews.
Walkie-talkie turn-taking
AI and human cut each other off.
Fails on interruptions, overlap and background speech
Doesn't listen, just transcribes
Fails to understand the person. Gets details wrong. Never asks to clarify.
"It has awkward pauses, awkward interruptions. If they cough in the middle of the agent's thing, it'll just shut up and be like, 'oh, are you still there?'"
pilot partner,
on their previous voice stack
why this happens
Voice agents are built on a fragmented stack.
It forces walkie-talkie turn-taking and loses information at every step.
This forces people to speak like AI, making every call frustrating.
metavoice
One model that listens, thinks and speaks in 350 ms.
It's learned the mechanics of conversation directly from human data.
Full duplex
Listens while it speaks, so interruptions, overlap and background speech don't break the call.
Reasons over speech, not text
Understands the words, how you said them and what's happening around you. Asks when it isn't sure.
Nothing to assemble
No ASR, TTS, turn detector or dialogue harness to build. You define the workflow, tools and personality.
Both channels stay open.
You barge in mid-sentence. It ducks, keeps listening, and picks right back up.
hear
Listen to a real call, unedited.
Jump to moments that break voice agents.
production
One model. Same control.
control
Guardrails
Filter or block a response before the caller hears it.
Debuggable
Inspect every layer of a turn: reasoning, tool calls and speech, in text and audio.
Observable
Streams into the evaluation and monitoring tools you already run.
yours, not a shared api
Build your agent
Set the workflow in a prompt, give it tools to call, choose its personality.
Your VPC
Deploy in your cloud. Data never leaves your network.
Improves over time
Coach it with feedback like a new rep. Or fine-tune it on your own calls.
Costs the same
At parity with your cascaded stack.
pilot
Run a paid pilot on one use case
for 30 days.
You choose the metric.
If we miss it, you don't pay.