Teaching machines
when to speak.
Speaking Machine Labs is a voice AI research lab. We build models for rooms full of people: models that hear every voice at once, follow the conversation, and know when to take the floor and when to give it back.
Above: a live spectrogram of six people and one machine in conversation. Drag to turn it. Point at or tap a voice to talk over whoever has the floor.
The gap
Give a language model the transcript of a group conversation and it follows the room 87.5% of the time. Put the best real-time voice agent in that same room and it gets 32.6%.
Text models against real-time voice agents on multi-party comprehension. MP-Bench, 2026.
The intelligence is already there. The timing is not. Voice AI was built for two people taking turns on a phone call. Real conversation has many voices, overlap, interruption, and silences that mean something. A machine that cannot tell a finished thought from a pause either talks over people or never speaks at all.
Three programs
Each program ends in a model, and each model is measured against people in real rooms. We will say more when the numbers are ready.
-
Underway
Floor
When should a machine speak?
A small model that listens to every speaker in a group and decides, many times a second, whether to speak, wait, keep the floor or give it up. It is designed to run on a CPU, with a decision latency of redacted.
-
Early research
Chorus
Can one model hear six people as six people?
A full-duplex voice model that hears each person as a separate stream, not one mixed signal. It listens while it talks and answers the person who asked. Architecture: redacted text.
-
Building
Agora
Can timing be learned before the meeting?
A gym of simulated meetings with many speakers, for training and scoring conversational timing, and for testing whether what a model learns in simulation still holds in a redacted.
Machines have spoken for 235 years.
In 1791, Wolfgang von Kempelen published the design of a machine that could speak. In 1939, Bell Labs filled a hall at the World's Fair with people who came to hear one. Machines learned to make speech long ago, and in the last few years they learned to make it sound human.
They still do not know when to speak. That is the part we work on.
How we work
Real rooms
We score every model on many-voiced meeting audio, with overlap, crosstalk and bad microphones. A clean two-person recording proves very little.
Small first
We use the smallest model that solves the problem. A decision made in milliseconds on a CPU beats a perfect one that arrives after the moment has passed.
Open measures
We intend to publish our benchmarks, so that anyone can measure the problem we care about, including the labs we compete with.
A small lab with a large room to work in.
Speaking Machine Labs is the research lab of MeetStream, the infrastructure that puts voice agents into Zoom, Google Meet and Microsoft Teams calls. We are hiring a few people in speech, full-duplex modeling and reinforcement learning.