Can Imagined Japanese Words Become a Voice? Inside a Prizewinning Brain-to-Voice Research Project

The promise is a voice—not a machine that reads minds

Speech can disappear while the desire to communicate remains. For people living with severe speech impairment, the prospect of converting deliberately imagined language into an audible voice is compelling. It is also one of the hardest problems in human-computer interaction: how can a machine reliably interpret an intention that makes no sound?

On October 9, Sojo University in Kumamoto announced that a joint research project with the Kyushu Institute of Technology had received an encouragement award at FIT2026, Japan’s 25th Forum on Information Technology. Its title translates as “Japanese Brain-to-Voice System Using EEG During Inner Speech.” The conference took place September 2–4; the announcement followed more than a month later.

The announcement establishes a research presentation and an award. It does not establish a commercially ready voice prosthesis, continuous Japanese speech generation or the ability to decode somebody’s private thoughts. Those limitations are as important to this story as the hope.

A student’s research crosses a university boundary

According to the university, the project linked the Ojima–Uemura laboratory in Sojo University’s Department of Information Science with Professor Kiyohisa Natsume’s laboratory at the Kyushu Institute of Technology. The work was presented as a poster by a researcher identified by the university as Kimura, a March 2026 Sojo graduate who is now a first-year master’s student at Kyushu Institute of Technology. The announcement does not provide Kimura’s complete name.

The named supervising researchers are Sojo professors Shuichi Ojima and Takumi Uemura, and Kyushu Institute of Technology professor Kiyohisa Natsume. Sojo’s IoT and AI Center supplied computing resources for model training. The recognition was the FIT Encouragement Award, associated with the September research presentation.

This is a story of collaborative academic development, not a company launching a medical product. The university lists a conference-proceedings paper but its announcement supplies no numerical decoding accuracy, sample size, vocabulary size, or patient outcomes. Without those details, the scale of the technical advance cannot be independently assessed.

What an EEG headset can—and cannot—see

Electroencephalography, or EEG, records electrical potential differences at sensors placed on the scalp. It avoids the surgery required by implanted cortical electrodes, and it offers rapid measurements. But each scalp sensor aggregates activity from broad neural populations after signals pass through tissue. Eye movement, muscle activity and electrical interference can complicate recordings.

Imagined speech adds a fundamental problem: there is no external sound to align with the brain data. In an experiment, a researcher may instruct a participant to imagine a particular syllable or word, but cannot directly observe the precise timing or content of that internal act. Even a successful experimental classifier might only distinguish two or three rehearsed choices.

Identifying one item from a fixed list, synthesizing a chosen syllable, producing natural continuous speech, and transcribing arbitrary private language are different tasks. Headlines that collapse them into “mind reading” exaggerate what research systems can do.

How the field reached this point

Brain-computer interfaces have been researched for decades as ways for users to control devices through recorded neural signals. Speech prosthesis research has accelerated as machine-learning systems have improved and neuroscientists have mapped more of the activity associated with speaking.

A 2024 review in Nature Reviews Neuroscience traced the shift from decoding speech-related activity in epilepsy monitoring to early feasibility work involving people with paralysis. It stressed the need to compare systems using meaningful metrics such as speed and error rates, rather than impressive demonstrations alone.

A separate 2025 Nature report described an implanted system synthesizing speech from the neural activity of a man with ALS as he attempted to speak. That result is an important milestone, but it used invasive electrodes and attempted speech. It cannot be used as evidence that Sojo’s scalp-EEG inner-speech system has achieved the same performance.

An instructive experiment in human learning

One particularly relevant independent study appeared in Communications Biology in February 2025. Fifteen healthy participants learned to control a two-choice imagined-syllable EEG interface over five consecutive days. On average, their control improved with feedback, although individual results varied considerably.

The finding points to an often overlooked issue: the person is not simply a passive supplier of neural data. The user may learn how to generate more consistently recognizable signals, while the decoding model also learns. A real communications device may depend on training both sides of that interaction.

The two-choice result did not establish open-vocabulary conversation, and it was not part of the Japanese university collaboration. It provides comparative scientific context, not a measure of this particular system’s achievement.

Why a Japanese-language system matters

Many widely publicized advances in speech neurotechnology have centered on English-language decoding. Japanese introduces its own engineering choices. Speech can be represented at levels including sounds, morae, syllables, words and grammatical units, while written output adds the complexities of kana and kanji.

A system constrained to several rehearsed Japanese syllables is different from one identifying words spoken silently, and more different still from one producing sentence-level audio in real time. The representation selected for model training affects data collection, accuracy measurement and the meaningfulness of any generated voice.

The university release does not specify which linguistic units its experiment used or describe a verified open-ended Japanese speech demonstration. Those remain reporting questions, not accomplishments that can responsibly be inferred from the project name.

Clinical promise has a demanding threshold

For a person who cannot speak because of ALS, stroke or another condition, a noninvasive interface might someday offer an additional communication option. The potential attraction is clear: sensors worn on the head rather than surgically implanted. But easier installation comes with noisier, less spatially precise signals.

A clinically useful system must work beyond a few controlled laboratory sessions. It must tolerate fatigue, changing electrode contact, movement and individual variation; allow corrections; protect sensitive data; and give the user control over when speech is generated. Performance measured on healthy volunteers cannot automatically be applied to patients.

A 2025 systematic review of imagined-speech BCI studies found that fewer than 6% reported real-time decoding. That provides a useful caution against conflating retrospective computer analysis with an everyday conversation device.

The privacy question should be framed accurately

The phrase “thought reading” creates both false hope and unnecessary fear. Current systems typically require a willing participant, sensors, a narrowly defined task and careful signal processing. There is no evidence in the Sojo announcement that arbitrary thoughts can be extracted remotely or without cooperation.

Still, the ethical questions are real. Who stores EEG recordings? How long do they remain identifiable? Could a system speak a mistaken interpretation aloud without the user’s approval? What rights would patients have to delete recordings or reject model outputs?

The best design principle is agency: the person, not the algorithm, should decide what counts as an intended message. Better decoders must be paired with consent, security and clear procedures for correcting errors.

Five answers that would turn an award into a breakthrough

Before judging clinical or commercial potential, researchers and readers need basic measurements: the number and characteristics of participants, the size and structure of the Japanese vocabulary, accuracy on unseen trials, real-time latency and performance with unfamiliar users. Published methodology and independent replication would make comparisons more meaningful.

The precise contribution of the model—and whether it generates speech directly or through an intermediate set of classified units—also matters. The university’s October 9 release identifies a conference paper, but does not provide these details. This article does not invent them.

The most interesting part of the story is not a futuristic claim that computers can finally hear thoughts. It is a more demanding possibility: carefully designed neural interfaces might eventually give some people a new way to express words they already intend to say. That is a goal worthy of attention, and of honest measurement.