OpenAI patents earbuds that read silently mouthed words
A new patent filing describes a wearable device that uses motion and strain sensors to translate jaw movements into digital commands.
The short version
- The system detects 'nonvocalized' speech, which includes words mouthed silently or spoken at levels below 30 dB.
- It uses a combination of motion sensors (accelerometers, gyroscopes) and shape-deformation sensors (strain gauges).
- The technology aims to feed silent inputs into large language models to perform tasks like querying databases or making calls.
- The patent describes a four-stage pipeline from raw sensor data to final command execution on devices like phones or robots.
Why it matters
This technology could allow users to control AI agents and smart devices privately without speaking aloud. It expands the way humans interact with voice assistants by making 'silent' speech a viable input method.
Detecting speech without sound
According to the filing, the device may detect speech events with volumes lower than 30 dB, 20 dB, or 10 dB. It is intended to capture the 'phonemic and articulatory intent' of a user. This means the hardware aims to understand what a person is trying to say based on the physical mechanics of their mouth, regardless of whether sound is produced.
How the sensors work
The system can also use a pressure sensor to monitor changes in an air-sealed cavity within the ear. By combining these different inputs, the device can distinguish between intentional speech and other movements like chewing, swallowing, or nodding. The application suggests using 'sensor fusion' to ensure the AI correctly identifies the user's intent from the complex stream of physical data.
From motion to machine learning
Finally, the system moves to an application interface where the identified words become commands. If the system detects a sequence of phonemes corresponding to a command like 'call XYZ,' it triggers a device-level instruction. The patent notes that the machine learning models can be tuned for a particular user, who would provide training data by mouthing a set of words while wearing the device.
Integration with AI agents
These commands can be sent to a wide range of remote devices. The patent mentions mobile phones, laptops, tablets, and even robotic assistants or home automation systems. It also suggests that the technology could be used in head-mounted displays or virtual and augmented reality devices to provide a hands-free, silent way to interact with digital environments.
Technical details and timeline
The filing appeared around the same time OpenAI acquired a hardware company called io in May 2025. While the patent does not explicitly mention the acquisition or specific inventors' roles, it highlights OpenAI's move into specialized hardware. The technology represents a potential shift toward more intimate and private ways for humans to communicate with AI agents in public or private spaces.
The patent filing structure
The document published by the USPTO is a comprehensive 21-page application. It includes a cover sheet, seven sheets of drawings, and 13 pages dedicated to the description and legal claims of the technology.
The filing provides a detailed roadmap of the invention. It covers the physical body of the wearable, the sensing systems for motion and deformation, and the processing systems required to turn those signals into words.
Filing timeline and regulations
The application followed a standard legal path for US patents. A provisional application was submitted in May 2025, followed by a nonprovisional application exactly 365 days later. This timing hit the final day of the one-year window allowed by federal law for claiming the benefits of a provisional filing.
The publication occurred 554 days after the initial provisional filing. This timeline aligns with the standard rule that the USPTO publishes applications roughly 18 months after the earliest priority date. The publication was released on a Thursday, which is the standard day for these updates.
Data classification and features
The application uses specific classification codes to define its scope. Three of the five codes are assigned to the subclass for speech analysis and recognition. The remaining two codes focus on hardware, specifically the measurement of surfaces and dimensions, as well as microphones and loudspeakers.
The system extracts complex data from the raw sensor signals to identify speech. Time-domain features include metrics like signal amplitude, root-mean-square energy, and waveform shape descriptors. It also analyzes frequency-domain features such as the spectral centroid and dominant frequency components to better classify the user's intent.
Key inventors and applicants
The patent application identifies OpenAI OpCo, LLC of San Francisco as the official applicant for the technology. The filing also names two specific inventors from the region: Douglas Stuart Brodie of San Jose and Daniel Boothe of San Francisco.
These individuals are credited with the development of the "In-Ear Silent-Speech Detection System." While the document does not specify their exact roles within the company, it establishes the formal origin of the hardware concept within OpenAI's corporate portfolio.
Visualizing the hardware design
The application includes several drawings to illustrate how the sensors interact with the human body. Specifically, Figure 2A and Figure 2B depict the sensing element as a zigzag trace that runs around the band of the ear tip. These illustrations help visualize how the hardware is positioned to capture movement.
Figures 3A and 3B show the tip flexing as the canal wall changes shape between two different positions. The document notes that these specific changes are exaggerated for the purposes of illustration. Additionally, the filing mentions that several elements could be arranged around the periphery of the compliant structure to capture deformation information at multiple points.
Neural network classifications
To translate physical movements into actual words, the system utilizes specific types of neural networks. The application lists abbreviations such as CNN, RNN, and TCN to describe the classifiers used in the processing pipeline. These models are responsible for mapping the extracted features to specific phonemes or full words in various languages.
The system handles the complexity of human speech by breaking it down into basic units. For example, the filing notes that a speaker can articulate forty-four possible phonemes in English. The model can therefore output a vector with forty-four entries, where each entry corresponds to the activation of a particular associated English-language phoneme.
Sources and links
- ReportingPPC Land: the article we sourced
First reported by PPC Land. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.