Imagine standing on a Tokyo subway platform, listening to the rapid-fire rhythm of Japanese around you. Then picture yourself in a pub in the UK, where English seems to move at a more leisurely pace. The two languages sound completely different. One is a quick run of simple syllables, the other a slower parade of more complex ones.
Yet they deliver information at almost exactly the same rate: about 39 bits per second. So does every other language that has been measured this way.
The finding comes from a study in Science Advances by Christophe Coupé, Yoon Mi Oh, Dan Dediu and François Pellegrino, “Different languages, similar encoding efficiency” (2019). Here’s what they found, as a set of questions and answers, with three figures to try, and what it means for AI.
What did the study find?
That human languages, however different they sound, carry information at a strikingly similar rate. Across 17 languages, the average information rate was 39.15 bits per second, with most speakers within about 5 bits per second of that figure (Coupé et al., 2019).
That’s remarkable because the languages themselves vary enormously. The authors note that Japanese and English differ by a ratio of about 1 to 11 in the number of distinct syllables they use, from a few hundred in Japanese to almost 7,000 in English. Some of the languages have no tones at all, Vietnamese has six. And yet the rate at which they get information across converges.
How did they measure it?
They recorded 170 native speakers (85 of them women) of 17 languages from 9 language families, across Europe and Asia. Each speaker read the same set of 15 short texts, translated into their own language, at their normal pace: about 240,000 syllables in all. The languages were Basque, Cantonese, Catalan, English, Finnish, French, German, Hungarian, Italian, Japanese, Korean, Mandarin, Serbian, Spanish, Thai, Turkish and Vietnamese.
Then they measured two things and multiplied them together.
Information density is how much information each syllable carries, in bits per syllable. It uses Shannon’s information theory, the mathematics behind digital communication. In this sense, a syllable carries more information when it is harder to predict from the one before it, which is easier in languages with many possible syllables. Density ranged from 4.8 bits per syllable for Basque to 8.0 for Vietnamese.
Speech rate is how fast people speak, in syllables per second. The average was 6.63 syllables per second, but individual speakers ranged from about 4.3 to 9.1. Japanese and Spanish speakers produced about 50% more syllables per second than Vietnamese and Thai speakers (CNRS via Phys.org).
Information rate = information density × speech rate. That product is the 39 bits per second.
Why do they all land on the same number?
Because density and speed trade off. Languages whose syllables carry little information are spoken quickly, and languages whose syllables carry a lot are spoken slowly. Move the slider and see how fast you’d need to speak to keep the information rate at 39 bits per second.
Dense and slow, or light and fast
The authors compare it to bird wings: big wings need few beats per second, little ones have to flap fast, and both birds fly (CNRS via Phys.org). Nobody decides this. It emerges over generations as languages change and their speakers adapt.
The study’s own phrase for it is that languages “stably inhabit an optimal range” of information rates, away from the extremes that individual speakers can still reach. The authors suggest the reason lies in how our brains process speech. Research on brain rhythms in the so-called theta range shows that the brain tracks speech at the pace of syllables, which may set a natural limit on how fast information can usefully arrive.
How fast is 39 bits a second?
Slow, by the standards of machines. A 1990s dial-up modem managed 14,400 bits a second. But it’s in the same league as other things humans do. A separate analysis by Jieyu Zheng and Markus Meister at Caltech estimated that human behaviour, from typing to solving a Rubik’s cube, runs at about 10 bits per second, while the senses take in something like a billion (Zheng and Meister, 2024).
Speech next to machines and senses
How fast do you speak?
Read this passage aloud at your normal pace, timing yourself with the buttons. It has 58 syllables.
Your speech rate
Some languages pack a lot into each sound and speak more slowly. Others use short, simple sounds and speak fast. Both get the same message across in about the same time. It is a trade between speed and detail, and each language seems to strike it.
What does “information” mean here?
Something narrower than it sounds. As the linguist Sean Trott points out in a careful review of the study, information in Shannon’s sense measures how predictable each syllable is. It says nothing about how much meaning the syllable carries. Two languages can be equally “dense” in this sense while saying quite different amounts. The study also covered 17 languages, all from Europe and Asia, with no African languages, so “universal” is a strong word for what is still a remarkable pattern.
The authors were careful too: they describe the result as “good-enough” control: a comfortable band that languages settle into.
What does it mean for AI?
People’s ears and brains didn’t change when machines learned to talk. So a few practical implications follow.
AI voices. A synthetic voice can speak at any speed, but listeners can only take in information at the human rate. Speeding up an AI voice to save time may simply lose people.
Translation and dubbing. When a sentence is translated, the number of syllables changes, sometimes a lot. A dubbed Japanese line and its English version can carry the same information in very different numbers of syllables, which is one reason dubbing and subtitling are harder than they look. AI dubbing tools have to manage this trade-off too.
Writing for people. AI can produce text far faster than anyone can read it. More output doesn’t mean more understanding. The bottleneck is still the human on the other end, so clarity and pacing matter more as the volume goes up.
Interfaces. Systems that adapt their speed and density to the listener, slower and simpler for complex or urgent information, would be working with human limits.
Where does this leave us?
With a pleasingly human result. Every language, however it sounds, has found its own route to the same destination: about 39 bits of information a second. AI can produce words at any rate it likes, but the people receiving them are still running at 39.
If you’re interested in how quickly the brain reads, rather than listens, see the ultra-fast reading machine inside your head.
Sources: Coupé, C., Oh, Y.M., Dediu, D. and Pellegrino, F. (2019) “Different languages, similar encoding efficiency: comparable information rates across the human communicative niche”, Science Advances, 5(9), eaaw2594; CNRS, “Faster and slower languages convey information at similar rates”, Phys.org, 5 September 2019; Trott, S. (2020) “Do different languages really convey information at the same rate?”; Zheng, J. and Meister, M. (2024) “The unbearable slowness of being: why do we live at 10 bits/s?”, Neuron; Wikipedia, Information theory.