Familiar

The duplex speech model for revenue calls. Formerly MetaVoice.

[Familiar @familiar_io]

MetaVoice is now Familiar.

Voice AI is growing fast, but most interactions still feel like a walkie-talkie. That's why voice agents today are largely limited to scripted, QA-style conversations.

For voice AI to move beyond that, it has to adapt to the way people already speak. Not the other way around.

Talking to AI should feel familiar, like talking to another person. That is the standard we named the company after.

At the core of Familiar is a full-duplex speech model built to make that possible. It listens while it speaks, for a continuous, two-way conversation.

More on the rename and where we're going: https://t.co/jfXi4ow1PN

[Siddharth Sharma @sidroopdaska]

I was an early engineer @wayve_ai , building end-to-end AI for self-driving.

The lesson: hand-built pipelines lose to models that learn the whole task.

Voice AI is at the same moment. Every voice agent today is a pipeline: transcribe → think → speak, with a separate component guessing when you're done talking. This turns the conversation into a walkie-talkie exchange, people get frustrated and hang up.

So at @metavoiceio we're building the other thing: one duplex model that listens while it speaks. Calls feel natural, like talking to a person, even through interruptions, overlaps and background voices.

I spoke with Theo @TheBreakAI about my journey, why we're building MetaVoice, how it differs from today’s turn-based solutions, and why revenue calls are the first use case.

https://t.co/ovbIM9stWa

[Familiar @familiar_io]

Mia & Leo are for developers building inbound and outbound voice agents.

We started with speech that sounds human. Mia & Leo are the next step: speech that can hold a human conversation.

Talk to them at https://t.co/2qWaZk8bhj

[Familiar @familiar_io]

The failures feel small. An awkward pause. A barge-in that’s missed. The agent talks over the caller. The caller corrects a date, but the agent misses it. A voice in the background derails the call.

But these compound quickly and the person gives up.

[Familiar @familiar_io]

A model is not production-ready because the demo sounds good.

Developers also need: Control over behavior. Failures they can debug. Calls they can observe. Costs that work at scale.

Mia & Leo are built for that.

[Familiar @familiar_io]

Voice agents have become much better at sounding human. But sounding human is not the same as having a human conversation. Teams building real-world agents told us that more than 40% of callers can hang up in the first 30 seconds.

[Familiar @familiar_io]

Today we’re launching Mia & Leo: AI personalities backed by purpose built duplex speech models for revenue calls.

It listens while it speaks so voice agents can handle interruptions, overlap, and background speech without turning the call into a walkie-talkie conversation.

[Siddharth Sharma @sidroopdaska]

Natural conversation maxxing!

A sneak peek of @metavoiceio's duplex speech AI.

Launching in 2 weeks.

[Familiar @familiar_io]

2/ That’s the future we’re building. And we’re fully leaning into the bitter lesson that scalable, data-driven methods win!

Read our blog on the first step towards that: https://t.co/h58WYvlDRj

[Familiar @familiar_io]

1/ Our belief @metavoiceio is that voice interactions with AI don’t go far enough.

What if talking to AI felt indistinguishable from talking to a person, even after years of daily interactions? Fluid back and forth, empathetic responses and a consistent personality.

[Familiar @familiar_io]

We’re also releasing a research preview of Speech-1.1, trained using ProsodyAlign, a novel post-training technique to more closely capture human-like speech patterns – pauses, stress, how the voice goes up/down in tone on different words. It's currently less stable than Speech-1.

[Familiar @familiar_io]

We��re releasing access to Speech-1, our conversational speech model designed specifically for customer phone calls (8khz telephony) to reduce call drop rates.

[Familiar @familiar_io]

Speech-1 fixes these problems!

[Familiar @familiar_io]

If you're in London, don't miss out! Reserve your spot at: https://t.co/rmuZTcdOfK. See you there! 👋

[Familiar @familiar_io]

@vatsal_aggarwal will be presenting at @AITinkerers in Shoreditch, London tomorrow! 🚀

Join us as he shares insights into MetaVoice's journey in building a 1B+ Speech+Text LLM. 🎙️

A huge shout out to @AITinkerers & @join_ef for organising this! 🙌