Narraloud Reading Room

Machines That Speak

Two centuries of trying to make text talk, from bellows and reeds to the narrators that fit in a pocket.

By Narraloud Editorial

The gargle before the voice

The first talking machines did not talk so much as gargle. Bellows pushed air through a reed, an operator worked leather flaps and levers, and audiences swore they heard words in the noise. What they mostly heard was hope. But hope is a fine instrument, and it kept the problem alive through a century in which inventors modeled the mouth in brass, in rubber, and at least once in painted wax.

Every one of those contraptions shared an assumption: to make speech, build a mouth. It was intuitive, it was heroic, and it was wrong in a productive way. The mouths got better and the speech did not, which is exactly the kind of failure that teaches.

Playing speech like an organ

The modern era began when the problem changed hands, from sculptors to engineers. The keyboard speech synthesizers demonstrated at the world's fairs of the late 1930s dropped the artificial mouth for an artificial larynx: a buzz, a hiss, and a bank of filters. Trained operators, almost all of them women hired for keyboard fluency, played sentences the way an organist plays a chorale, and after months of practice could make the cabinet say almost anything on request.

The crowds queued for the novelty, but the fairs proved something quieter and more durable: speech could be assembled from parts that no mouth contains. Once that was true, the rest was a schedule.

Each generation of talking machines succeeded exactly where it stopped imitating the mouth and started imitating the ear.

Rulebooks, recordings, and learning

Mid-century systems traded the operator for rules. If speech was filters, then writing down which filters, and when, became a writing system of its own. For three decades the state of the art was a growing rulebook and a shrinking cabinet. The voices were tireless, intelligible, and unmistakably synthetic, and a generation of blind and low-vision listeners came to prefer them at speeds no human announcer could match. That preference is worth sitting with: the measure of a voice is the listener, not the lab.

Then the statistical turn replaced the rulebook with recordings, and the neural turn replaced the recordings with models of them. Each step moved the work from explaining speech to learning it. The rulebooks did not survive. The insight did.

The voice in your pocket

The newest chapter is architectural rather than acoustic. A neural voice small enough to run on a phone changes the terms of the whole two-century project: the machine that speaks is finally the one in the reader's hand, and it works in the places reading actually happens, on a plane, underground, past the end of the signal. The fairground crowd of 1939 lined up to hear a cabinet say good afternoon. Their great-grandchildren expect the afternoon read to them on the train, at their own speed, in a voice they chose.

Nothing about this was inevitable. The talking machine was kept alive by operators who practiced for months, by listeners who tolerated the gargle for the sake of the words, and by a stubborn conviction that text and voice are the same thing at different speeds. The devices changed. The conviction did not, and it is the reason this page can read itself to you.