Subvocalization

When you read this sentence, you probably hear the words in your head. That’s exactly what’s called subvocalization: silently speaking along while reading or thinking. It doesn’t stay purely in your head. Your tongue, larynx, and lips move along minimally, often by mere fractions of a millimeter. From the outside you can’t see anything, but sensitive measuring devices can capture these movements. The term originally comes from reading psychology, but is now also used in technology.

Between reading speed and silent control

In school, subvocalization usually comes up as a supposed problem. Anyone who silently speaks along with every word can read at most as fast as they could speak. That’s roughly 150 to 250 words per minute. For decades, speed-reading courses have therefore promised that you need to train yourself out of this inner speech.

Research takes a more nuanced view. Inner speech helps with understanding difficult sentences and with retaining content. If you suppress it completely, comprehension of demanding texts declines. So a common misconception is that subvocalization is generally bad. With simple texts it slows you down, with complex ones it supports you.

The effect became technically interesting for a completely different reason. If measurable signals arise during silent speech, those signals can be read out. This could be used to control a device without making a sound. This is especially relevant for loud environments and for people without a voice.

What sensors on the neck and jaw measure

The most common method is called electromyography, or EMG for short. Electrodes are attached to the skin and measure the weak electrical voltages that muscles generate when they contract. On the jaw, under the chin, and on the neck lie exactly the muscles that work during speech. They also fire during silent inner speech, just much more weakly.

On their own, these readings are unusable. They look like jittery noise and vary greatly from person to person. This is where an AI model comes in — a program that learns patterns from many examples. A test subject silently mouths certain words thousands of times, and the corresponding signals are recorded. The model then learns which pattern belongs to which word.

A well-known research project on this is AlterEgo from MIT, presented in 2018. The device sits on the jaw and, in initial tests, recognized a limited vocabulary with about 90 percent accuracy. It’s important to distinguish this from mind reading: what’s measured are muscles, not brain activity. Anyone who merely thinks without silently speaking produces no usable signal.

From reading exercises to silent assistants

In everyday life, you’re most likely to encounter the term when learning. Reading trainings and learning apps recommend reducing inner speech when skimming texts. For vocabulary or formulas, the same guides advise the opposite, because speaking along supports memory.

In tech news, subvocalization comes up in the context of wearables, meaning devices worn on the body. Several companies are working on headphones and neckbands that allow voice assistants to be operated silently. The appeal is obvious: you could give a command on the train or in an open-plan office without disturbing anyone.

A second field of application is medicine. People who can no longer speak after laryngeal surgery often retain muscle control. For them, such a system would be a substitute for their lost voice. At the same time, there’s an open privacy question. A device that listens in on silent speech raises different problems than a microphone you can simply switch off.