Speaking Machine Labs

Origins

The lab takes its name from the first machine that could speak. For 235 years the work was about making speech. Now we live in the era of natural voice, and the open problem is timing.

  1. 1791

    Detail of an engraved plate of Kempelen's speaking machine, showing a box and folded bellows.
    From Kempelen's book on the mechanism of human speech, Vienna, 1791.

    Kempelen's speaking machine

    Wolfgang von Kempelen spent about twenty years on a machine that could talk. It had bellows for lungs, a reed for a voice, and a soft mouth that the operator shaped with one hand. It could say words and short phrases.

    In 1791 he published everything: the theory of how people make speech, and drawings of every part, so that others could build the machine again. People still do.

  2. 1846

    Engraving of Faber's Euphonia: a mannequin head on a frame above a large keyboard cabinet.
    Joseph Faber's Euphonia, about 1846.

    Faber's Euphonia

    Joseph Faber built a talking head that an operator played from a keyboard, with bellows below. It was shown in London in 1846. It could speak, and it could sing.

    Audiences found it uncanny. A machine that speaks is not yet a machine you can talk to.

  3. 1939

    A crowd in a curved hall under a mural titled The Voder watches an operator at a console.
    The Voder at the New York World's Fair. Bell System Technical Journal, 1940.

    The Voder

    Homer Dudley's team at Bell Labs built the Voder, which made speech from electronic signals. Trained operators played it like an instrument, with keys, a wrist bar and a foot pedal for pitch.

    At the World's Fair, crowds filled the hall to hear it. It was a room full of people listening to one machine, which is close to the problem we work on today, turned around.

  4. 2016

    A modern wooden reconstruction of Kempelen's speaking machine: a box with holes on top and black bellows.
    A modern reconstruction of Kempelen's machine.

    Voices that sound human

    Neural speech synthesis made machine voices sound like people for the first time. Within a few years, a synthetic voice on a phone call could pass for a person.

    The machines could speak well. They still waited for a full silence, then talked, like a walkie-talkie.

  5. 2026

    A natural voice, turned. From the side it is a waveform. Turn it, by scrolling or dragging, and it is a spiral: loudness and pitch at once.

    The era of natural voice

    Today the frontier is natural voice. In October 2026, Block, Jack Dorsey's company, showed Dessa: a full-duplex model that goes from speech to speech, trained only on recorded human conversation, with no text in the loop. It listens while it talks, and Block says it can laugh like a person. It is a research preview.

    A machine can now sound like a person in a conversation with one person. Put it in a meeting with six and it falls apart. It cannot tell who is speaking to whom, and it does not know when it is its turn.

  6. Next

    Machines that know when to speak

    Speaking Machine Labs works on that last step: voice models that hear a group as a group and take the floor only when it is theirs to take.

    Our research direction

The name

Speaking machine, singular, after Kempelen's. It is a reminder that this problem is older than computers, and that the first person to work on it published his drawings so that others could build on them.

Speaking is only half of what our models do. The other half is listening, and in time, seeing. The name stays short; the work is about machines that listen, see and speak.

Sources

All archival images are from Wikimedia Commons and are in the public domain. We toned them in black and white. The 3D figures on this site are drawn live in your browser.