A short history of talking machines

Updated October 31, 2026

People have tried to build machines that talk for more than two hundred years. The story runs from bellows and leather tubes to the neural voices you can use today.

Mechanical speaking machines

In the late 1700s, inventors built mechanical models of the human voice. The best known is Wolfgang von Kempelen's speaking machine, described in his 1791 book on the mechanism of human speech. It used a bellows for lungs, a reed for the vocal cords and a leather tube shaped by hand for the mouth, and could produce some vowels, consonants and short words. It proved an important idea: speech can be made by imitating how the body produces it.

The Voder: electronic speech

At the 1939 New York World's Fair, Bell Labs demonstrated the Voder, designed by Homer Dudley. It made speech electronically from a buzz and a hiss, shaped by filters. A trained operator played it with keys and a foot pedal, a bit like an organ. It took many months of practice to make it speak clearly, but audiences heard a machine talk for the first time.

Computers start to talk and sing

In 1961, researchers at Bell Labs used an IBM 704 computer to sing "Daisy Bell" ("Daisy, Daisy, give me your answer do"), one of the earliest demonstrations of computer speech synthesis. Through the 1960s and 70s researchers built the first systems that could convert ordinary written English into speech automatically, including work at Japan's Electrotechnical Laboratory and at MIT.

Reading machines and talking toys

In 1976 the Kurzweil Reading Machine combined a scanner, character recognition and speech synthesis so that blind people could have printed books read aloud. It was one of the first uses of text to speech for accessibility, the field that still drives much of its development.

In 1978 Texas Instruments released the Speak & Spell, a children's toy that used a single speech chip. For many people it was the first talking electronic device they owned.

DECtalk and a famous voice

In the 1980s, formant synthesizers based on the research of Dennis Klatt became widely used, most famously in DECtalk. A voice from this family of synthesizers became the voice of physicist Stephen Hawking, who kept it for decades even when more natural options appeared, because he had come to think of it as his own.

Recorded voices stitched together

In the 1990s and 2000s, concatenative systems became popular. They recorded a real speaker reading hours of material, cut it into small units, and joined the right pieces to form new sentences. The best systems sounded quite natural, and they powered GPS devices, phone menus and early screen readers.

The neural era

In 2016 DeepMind introduced WaveNet, a neural network that generated speech audio sample by sample and sounded strikingly more human than earlier systems. Google's Tacotron and many other models followed. Today's neural voices learn pronunciation, rhythm and tone from recorded speech, and run fast enough to work on ordinary servers and even phones.

Open-source projects have brought this quality to everyone. The natural voices on TTSMP3 come from Piper and Kokoro, two open neural speech engines, alongside eSpeak NG, a formant synthesizer that covers many more languages.

What hasn't changed

From the 18th century to now, the goal has stayed the same: make the sounds of human speech, and make them clear enough to understand. And from the Kurzweil Reading Machine to today's listen buttons, one of the most important uses has always been giving people access to text they could not otherwise read.