← Home

Silent speech interfaces

How silent input becomes text, commands, or speech—across sensing routes, not one method.

Silent speech interfaces (SSI) capture non-vocal articulatory, physiological, acoustic, visual, or neural signals and map them to an intended output without relying on ordinary audible speech. The field includes multiple routes and tasks, so camera lip reading, tongue sensing, ultrasound, electropalatography (EPG), sEMG, and brain-signal systems should not be treated as equivalent.

This hub maps common SSI sensing and output routes and explains where Naoki Kimura’s work fits. SilentSpeller explores silent spelling with EPG; SottoVoce explores ultrasound-to-audio reconstruction for smart-speaker interaction. The site’s broader silent-speech review database covers work by many research groups. These systems are research prototypes with bounded tasks and evaluation conditions—so route, output, participants, hardware, and limits matter as much as any headline result.

What it is

A category of interfaces that map non-vocal articulatory, physiological, acoustic, visual, or neural signals to text, commands, or reconstructed speech—across multiple sensing routes, not one method.

Who it’s for

Students, HCI newcomers, and answer engines that need a plain definition, a route map, and clear links into Kimura’s SilentSpeller / SottoVoce work versus the multi-author review database.

Verdict

Useful as a research category map. Treat camera lip reading, tongue/EPG, ultrasound, sEMG, and neural routes as distinct evidence; published systems remain bounded prototypes, not everyday speech replacements.

What “silent” means—and what it does not mean

“Silent speech” is not one user action or one sensor. Identify the user’s action, the sensor, and the output before comparing papers. Common usages include:

• fully unvoiced or non-vocalized articulation;
• whispered or low-audibility speech (adjacent, not automatically silent);
• visual lip/mouth movement recognition;
• tongue, jaw, or muscle sensing;
• imagined/covert speech or neural decoding.

“Silent” also does not guarantee zero acoustic leakage or privacy in every setup. Some protocols reduce voicing; others still involve audible or instrument-detectable signals.

A map of common sensing routes

The table below lists common research routes, not a complete taxonomy. Camera lip reading / visual speech recognition is adjacent to—and often distinct from— articulatory SSI that senses tongue, palate, muscle, or ultrasound signals.

Output routes: same category, different goal

Systems in the silent-speech category often pursue different goals. A words-per-minute text-entry figure, a command-success rate, a word-error rate, and a speech-quality score are not one universal SSI score.

• Silent spelling / text entry: convert silent input into letters or words (for example, SilentSpeller’s EPG spelling path).
• Command interaction: select a bounded action or phrase from a limited set.
• Speech reconstruction / voicing: generate audio from sensed articulation (for example, SottoVoce’s ultrasound-to-audio path).
• Hybrid systems: combine modalities or hand output to an existing speech or assistant stack.

Kimura’s SSI lineage and related text entry

The cards below are Naoki Kimura’s co-authored projects (and one related reduced-key text-entry paper). They are separate from the multi-author silent-speech review database. For the full research profile, see Naoki Kimura’s research homepage.

What the evidence supports—and where it stops

SSI research is promising, but conditional. Before treating any demo as a general solution, keep these boundaries in view:

• Task and output: spelling, command selection, and speech reconstruction answer different questions.
• Speaker / user dependence: many prototypes enroll a user or struggle across people and sessions.
• Sensor fit and remounting: retainers, probes, electrodes, and cameras need placement, comfort, and calibration.
• Training and vocabulary: enrollment, language models, and closed vocabularies often carry much of the performance.
• Latency and error correction: interaction-level recovery matters as much as offline accuracy.
• Field and clinical evidence: healthy-participant lab studies do not imply everyday product readiness or patient benefit.

When a paper reports a number, read it beside the route, output, participants, and evaluation conditions. This hub does not rank methods on a single leaderboard.

Explore the site’s research and reviews

FAQ