OpenAI has a pending US patent application for an earbud-style device that works out which words its wearer is mouthing without making a sound, by measuring how the ear canal moves and changes shape as the jaw works. The application, US 2026/0369273 A1, titled "In-Ear Silent-Speech Detection System," was published on November 19, 2026. It names OpenAI OpCo, LLC of San Francisco as the applicant and two inventors, Douglas Stuart Brodie of San Jose and Daniel Boothe of San Francisco.

In Short

OpenAI has asked for a US patent on an earbud that reads words you mouth silently, using motion and strain sensors that pick up how your jaw moves the inside of your ear canal. It matters to anyone who reaches people through voice assistants and AI agents, because the filing sends those silent words to language models that can compose messages, update schedules, query databases or start a phone call. Nothing has been granted or launched, and the document contains no accuracy figures, no product name and no mention of advertising, consent or data retention.

What the filing covers

The document runs to 21 pages: a cover sheet, seven sheets of drawings and 13 pages of description and claims. The abstract states the idea in one sentence: "A wearable nonvocalized speech detection system may include a body configured to be positioned at least partially in an ear canal of a user, a sensing system positioned in the body and configured to detect motion of the ear canal resulting from articulation of a mouth of the user corresponding to a nonvocalized expression of one or more words and generate motion information representative of the detected motion, and a processing system operatively coupled to the sensing system and configured to generate, based on the motion information, a representation of the one or more words."

The paper trail is regular. Provisional application 63/805,864 was filed on May 14, 2025, under the same title. The full nonprovisional application, 19/677,432, followed on May 14, 2026, exactly 365 days later and on the last day of the one-year window that 35 U.S.C. 119(e) allows for claiming a provisional's benefit. Publication came 554 days after the provisional filing, by PPC Land's calculation, in line with the rule that US applications are published about 18 months after their earliest priority date; the USPTO publishes on Thursdays. A published application confers no patent rights; the claims can still be narrowed, rewritten or rejected during examination.

The provisional date has a neighbor in OpenAI's corporate history. OpenAI acquired io, a hardware company co-founded in 2024, in May 2025 for approximately $6.5 billion, the same month the provisional was filed. The application itself never mentions io, Jony Ive, ChatGPT or any product name, and it does not state the inventors' roles.

Field Entry
Publication number US 2026/0369273 A1
Publication date November 19, 2026
Title In-Ear Silent-Speech Detection System
Applicant OpenAI OpCo, LLC, San Francisco, CA
Inventors Douglas Stuart Brodie, San Jose, CA; Daniel Boothe, San Francisco, CA
Application number 19/677,432
Filed May 14, 2026
Provisional application 63/805,864, filed May 14, 2025
International classification G10L 15/24, G01B 7/28, G10L 15/26, G10L 15/30, H04R 1/08
CPC classification G10L 15/24, G01B 7/28, G10L 15/26, G10L 15/30, H04R 1/083
Claims 20 (independent claims 1 and 10)
Drawings 7 sheets, figures 1, 2A, 2B, 3A, 3B, 4, 5, 6

Three of the five classification codes sit in G10L, the subclass for speech analysis and recognition. The other two point at the hardware: G01B covers the measurement of dimensions, surfaces and contours, and H04R covers microphones, loudspeakers and similar transducers.

What counts as silent speech

The application defines its subject broadly. According to paragraph 24, a nonvocalized expression of words "occurs when a person produces biomechanical movement of one or more muscles, joints, or other anatomical structures of their face, including the temporomandibular joint." A user may "mouth" or "pantomime" words, which "may result in biomechanical motion without sound, or with low sound output (e.g., below a whisper level)."

Silence is not required. The same paragraph says nonvocalized expressions "may be inaudible to other listeners, but need not be completely silent or completely inaudible." Paragraph 28 gives the loudness bands: the sensing system may detect speech events "having a volume less than about 30 dB, less than about 20 dB, less than about 10 dB, or lower." Claim 20 carries the 30 dB figure into the claims, covering a microphone that picks up phonated speech "below about 30 dB."

Then the definition reaches further. "For example, in some cases of subvocalization or silent speech, miniscule movements of a person's anatomy may occur while the person is thinking about or reading words (even if those movements are not visibly perceptible to observers or even perceptible to the person)," according to paragraph 24. "In some cases, the ear-worn sensing devices described herein may detect nonvocalized expressions using such miniscule and/or unconscious movements and/or biomechanical responses." The claims themselves speak of a "nonvocalized expression of one or more words" and do not use the word unconscious, but the description stretches the concept to movements the wearer may not know they are making.

The background section frames the problem with a plain observation: "Electronic devices often receive and/or sense inputs from users. In some cases, inputs may be vocalized inputs, such as a user audibly speaking words." The answer the application offers is to read the jaw instead. According to paragraph 22, the devices "may leverage biomechanical sensing systems to detect speech-related jaw movement, specifically via the temporomandibular joint (TMJ)." The paragraph continues: "By capturing fine movements of the jaw, the ear-worn sensing devices can determine phonemic and articulatory intent (e.g., the nonvocalized words), even in the absence of sound."

Two ways of reading the ear canal

The two independent claims split the sensing problem in two.

Motion

Claim 1 rests on motion. A body sits at least partly in the ear canal, and a sensing system inside it detects "motion of the ear canal resulting from articulation of a mouth of the user." Paragraph 29 lists the hardware as "an inertial measurement unit, an accelerometer, a gyroscope, a magnetometer, and/or various combinations of these." One example "includes a 3-axis accelerometer and a 3-axis gyroscope," operating at "a configurable sample rate, which may range between about 100 Hz to about 1000 Hz." The parameters it can track are listed as "translation, rotation, acceleration, velocity, orientation, and the like." Dependent claims 2 and 3 name the accelerometer and the inertial measurement unit.

Shape

Claim 10 rests on deformation. A compliant structure, such as a soft tip or gasket, presses against the ear canal wall, and a sensing element coupled to it detects "a change in shape of the surface of the ear canal." According to paragraph 37, the compliant structure may be made from "a flexible polymer material (e.g., silicone, thermoplastic polyurethane, polymer foam, etc.)" and is "biased into contact with the surface of the ear canal." The sensing element may be "insert molded with the compliant material," or attached "via adhesive, fasteners, ultrasonic welding, or the like."

Paragraph 42 lists the candidate elements: "a strain gauge, a piezoresistive element, a piezoelectric element, a quantum tunneling material, a capacitive displacement sensor, an optical sensor, or the like." Claims 5 and 11 single out the strain gauge. FIG. 2A and FIG. 2B draw the element as a zigzag trace running around the band of the ear tip, and FIGS. 3A and 3B show the tip flexing as the canal wall changes shape between two positions, a change the application says "is exaggerated for the purposes of illustration." Paragraph 40 adds that several elements could be arranged "around a periphery of the compliant structure in order to capture deformation information at a plurality of points or locations around the ear canal."

The seal does double duty. According to paragraph 27, the compliant structures "may also act as a seal that defines an at least partially air-sealed cavity in the user's ear canal," which a pressure sensor can then monitor for changes as the mouth and tongue move. The body itself "may be customized to match a particular user's ear canal (e.g., by creating the body portion from a mold of a user's ear canal)," according to paragraph 26.

The other sensors on the list

Beyond the two claimed approaches, the description catalogues a long list of sensors that could sit in or near the ear. FIG. 4 names four subsystems: motion (408), deformation (410), vibration (412) and pressure (413). Paragraph 61 goes much further.

Sensor type named in the application What the application says it can detect
Inertial measurement units (accelerometers, gyroscopes, magnetometers) Movement, orientation or tremors of the head or jaw
Strain gauges, piezoresistive and piezoelectric elements, quantum tunneling material, capacitive displacement and optical sensors Deformation of the device housing or the ear canal
Contact or bone conduction microphones Vocalizations, chewing, swallowing, speaking, tongue movement, air passing through the vocal tract
Air pressure sensors Pressure changes in the sealed ear canal from jaw and tongue movement, sinuses, eustachian tubes
Electromyography electrodes on the skin of the inner ear Myoelectric activity of facial or neck muscles
Proximity sensors (infrared, VCSEL-based, ultrasonic) Deformation or motion within the ear canal or surrounding tissue
Photoplethysmography sensors Local blood volume changes for heart rate or oxygen saturation
Colorimeters Changes in local blood volume or pressure
Thermistors and other thermal sensors Local skin temperature, changes due to blood flow or airflow
Capacitive or resistive sensors Contact, skin impedance, hydration levels
Acoustic sensors Otoacoustic emissions, tympanic membrane vibration
Ultrasonic transducers Echo-based detection of anatomical movement or fluid levels
Ambient light sensors Lighting context, light-based signals through the ear canal

Sensor fusion ties these together. According to paragraph 64, data from an inertial measurement unit "may be fused with data from strain gauges and electromyography sensors to distinguish between head movements caused by intentional gestures (e.g., nodding) and those caused by jaw activity (e.g., chewing, swallowing, or speaking)." Paragraph 71 gives a concrete case: a jaw opening "may be inferred and labeled on receiving and/or sampling a pattern of upward accelerometer movement combined with a bone conduction signal from a microphone registering a pop or click of the TMJ."

How the four-stage pipeline works

FIG. 5 lays out a signal processing pipeline, numbered 500, in four stages. According to the application, it can run on the earbud, on a connected device, on several devices, or split between them, and the worked example pairs the earbud with "a cellular phone." The document describes the stages as follows.

  1. Sensing (stage 502). The earbud, or a pair worn one in each ear, polls its sensors for motion, shape changes, air pressure and vibration. From those readings it can "determine or infer a position or motion of the wearer's TMJ," infer whether the wearer "has constricted or relaxed one or more muscles," whether the wearer "has swallowed or has engaged the tensor veli palatini, the levator veli palatini, and/or salpingopharyngeus," and whether the head has nodded or shaken. The output is a stream of digital values.
  2. Conditioning and feature extraction (stage 504). The raw or fused signals are cleaned up with filtering ("low-pass, high-pass, band-pass, or notch filtering"), baseline drift removal, artifact suppression, gain normalization, resampling, synchronization across sensor types and timestamping. Features are then extracted. Time-domain features "may include signal amplitude, root-mean-square energy, zero-crossing rate, signal envelope, autocorrelation metrics, or waveform shape descriptors"; frequency-domain features "may include dominant frequency components, spectral centroid, bandwidth, and/or harmonic structure." The results can be packed into feature vectors.
  3. Classification (stage 506). Neural networks or other classifiers map the features to phonemes or words. FIG. 5 lists "CNN/RNN/TCN," abbreviations that usually stand for convolutional, recurrent and temporal convolutional networks, though the application does not spell them out. Paragraph 75 gives the arithmetic for English: "in English, forty four possible phonemes may be articulated by a speaker," so the model can output "a vector with forty four entries, each of which contains a value corresponding to an activation of a particular associated English-language phoneme." If the wearer makes the anatomical motion for an "O" sound, the "O" entry lights up. The stage can also classify whole words, "in one or more languages."
  4. Application interface (stage 508). The phonemes or words become actions. FIG. 5 lists "display output," "text to speech" and "command execution." According to paragraph 78, "if a sequence of phonemes is classified into a spoken command such as 'call XYZ,' the fourth stage 508 may interpret the detected sequence as a device-level instruction to initiate a phone call."

Training is personal in at least one claimed form. Claims 9 and 16 cover a machine learning model "trained on training data including sample motion information" (or deformation information) "generated by the sensing system in response to nonvocalized expressions by the user of a set of training words." Paragraph 76 adds that a model "may be tuned for a particular user," who would mouth a set of training words or phonemes while wearing the device. In practice, that points to an enrollment step before the system recognizes a given wearer, though the document does not describe one as a product flow.

From mouthed words to agents

The fourth stage is where the filing connects to OpenAI's core business. Paragraph 79 says it "may include natural language processing modules configured to parse grammatical structure (in one or more languages), perform named entity recognition, resolve ambiguities (e.g., pronoun resolution), or derive semantic intent from input sequences." It continues: "Such operations can be performed by and/or with one or more large language models or generative output engines prompted to extract user intents from structured data including the plain text or phoneme output from the third stage 506." The device may also keep "dialog state tracking to maintain context across multiple user interactions or pantomimed words."

Paragraph 80 extends the list to agentic systems "capable of producing multimodal outputs or triggering sequences of actions based on inferred user goals." The tasks named are specific: "composing messages, updating schedules, querying databases, generating documents, retrieving media, or navigating user interfaces via multimodal or voice-based interactions." Real-time output could include "closed captioning, real-time speech-to-text transcription, language translation, or the like."

The list of devices that might receive those commands is equally wide. According to paragraph 54, the remote device "may be any suitable device, including, without limitation, a mobile phone, a laptop or desktop computer, a tablet computer, an emotive companion, a robotic assistant, a television, an appliance, a home automation system or component, a lighting controller, or the like." Paragraph 52 adds head-mounted displays and virtual- or augmented-reality devices.

Where the processing happens

The application leaves the split between earbud and phone open. Paragraph 65 says sensor fusion can run "locally on the ear-worn device itself using embedded processing hardware, such as a microcontroller or digital signal processor," or raw or pre-processed data can be streamed to a smartphone, smartwatch or edge computing device for "more computationally demanding tasks such as machine inference, filtering, clipping correction, estimation via Kalman filtering, or real-time decision trees." The deciding factors are named: "The division of processing between devices may be configurable based on available processing resources, battery life, latency, or user preference."

Streaming "may be wireless (e.g., via Bluetooth, Ultrawide Band, Wi-Fi, cellular, or the like)," according to paragraph 57. A pair of earbuds can divide the work, with one carrying the sensing system and a communication system and the other a processing system and a communication system. Paragraph 35 keeps conventional functions in scope: speakers for "music, voice communications, audiobooks, or the like," and outward-facing microphones for calls, voice commands and speech-to-text.

Conversations nobody overhears

One use case stands out. Paragraph 53 describes two wearers talking through their devices: words "silently spoken by a first user may be sent to a second user's ear-worn sensing device to be output as a spoken output," and the reply travels back the same way. "Since the ear-worn sensing devices may provide audio outputs that are only audible to the wearer (as earbuds are generally largely silent to bystanders), a communication system using the wearable nonvocalized speech detection system described herein may facilitate largely silent communications. This may be useful to ensure communication privacy and allow communications in environments where speaking aloud would be impractical." Claim 8 covers the core of it: a remote device that generates "an audio output of the one or more words."

That sentence about communication privacy is the only place the document uses the word privacy. It contains nothing on consent, encryption, storage periods or who can access the sensor data.

The claims

The 20 claims split into two families. Claim 1 reads in full: "A wearable nonvocalized speech detection system comprising: a body configured to be positioned at least partially in an ear canal of a user; a sensing system positioned in the body and configured to: detect motion of the ear canal resulting from articulation of a mouth of the user corresponding to a nonvocalized expression of one or more words; and generate motion information representative of the detected motion; and a processing system operatively coupled to the sensing system and configured to generate, based on the motion information, a representation of the one or more words."

Claim 10 swaps motion for shape. It requires a body, "a compliant structure coupled to the body and configured to contact a surface of the ear canal," and "a sensing system comprising a sensing element coupled to the compliant structure," configured to "detect, via the sensing element, a change in shape of the surface of the ear canal resulting from articulation of a mouth of the user corresponding to a nonvocalized expression of one or more words," with the processing system generating the words "based on the deformation information."

Claim Depends on What it adds
1 Independent In-ear body, motion sensing of the ear canal during mouthed words, processing into a representation of the words
2 1 Accelerometer
3 1 Inertial measurement unit
4 1 Compliant structure touching the ear canal, sensing element detecting shape change, words also based on deformation
5 4 Sensing element includes a strain gauge
6 1 Motion results from movement of the temporomandibular joint
7 1 Wireless link sending the words to a remote device
8 7 Remote device generates an audio output of the words
9 1 Machine learning model trained on the user's own mouthed training words (motion data)
10 Independent In-ear body, compliant structure, sensing element detecting ear canal shape change, processing into words
11 10 Sensing element is a strain gauge
12 11 Strain gauge at least partially embedded in the compliant structure
13 10 Compliant structure comprises a flexible polymer
14 10 Adds motion sensing; words also based on motion
15 10 Shape change results from temporomandibular joint movement
16 10 Machine learning model trained on the user's own mouthed training words (deformation data)
17 10 Wireless link sending the words to a remote device
18 10 Microphone detecting vibrations from phonated speech
19 18 Microphone is a bone conduction microphone
20 18 Phonated speech below about 30 dB

None of the claims mentions language models, agents, electromyography or the health sensors. Those appear only in the description, which means they are disclosed but not, at this stage, claimed.

The drawings, and an emotive companion

Figure Sheet What it shows
FIG. 1 1 of 7 Device 101 with a stem, body 102, sensing system 104, beside ear canal 106
FIG. 2A 2 of 7 Device 201 with a rounded body 202, compliant structure 204 and deformation sensing element 206, beside ear canal 208
FIG. 2B 3 of 7 Device 201 on its own (also on the cover page)
FIGS. 3A, 3B 4 of 7 Compliant structure 204 and element 206 flexing as ear canal surface 300 changes shape
FIG. 4 5 of 7 System 400: sensing system 402 (motion 408, deformation 410, vibration 412, pressure 413), processing system 404, communication system 406
FIG. 5 6 of 7 Pipeline 500: sensing 502, processing 504, "ML interference engine" 506, application interface 508
FIG. 6 7 of 7 Electronic device 600 with processing unit, memory, touch and force sensors, haptic device, battery, display, audio input and output, positioning system, sensors, image capture system

FIG. 1 resembles a stem-style earbud; FIG. 2A shows a stemless, bulb-shaped design. FIG. 6 is the odd one out. Paragraph 20 describes it as "a system diagram of the components of an example emotive companion," and paragraph 89 says the displays "may convey information about the attention state of an emotive companion, the mode of operation of an emotive companion, or other information." The application never defines an emotive companion, and paragraph 81 describes the same figure as a generic electronic device that can represent the earbud, a phone, a base unit or an AR or VR headset.

A few drafting slips are visible in the published text. FIG. 5 labels stage 506 an "ML INTERFERENCE ENGINE," where the text treats it as a classification stage, which makes inference the likely intended word. Paragraph 74 lists "bone condition data" among model inputs, beside microphone data, where the rest of the document speaks of bone conduction. Paragraph 44 repeats paragraph 34 word for word. None of these changes the substance of the claims.

OpenAI's hardware push so far

The filing lands in the middle of a well-documented hardware push. In a podcast interview on July 23, 2025, about two months after OpenAI bought io, Sam Altman described his interest in a new kind of computer built around AI that assists people continuously. A week later, on July 29, 2025, Munich audio company Bragi announced a collaboration with OpenAI to put OpenAI voice interfaces into headphones, activated by a single press or a hotword. The patent application describes the opposite of a hotword: an input channel that needs no audible trigger at all.

The push has also drawn litigation. On July 10, 2026, Apple sued OpenAI, io Products and two former Apple employees in federal court in San Jose over alleged trade secret theft. One defendant, Tang Yew Tan, is OpenAI's Chief Hardware Officer. Apple's complaint says OpenAI employs or interacts with more than 400 former Apple employees, and PPC Land's coverage of the case noted that OpenAI had confirmed its first hardware prototypes by November 2025 and that OpenAI and Jony Ive aim to bring a first device to market. Neither inventor on the silent-speech application is named in that suit, and the application does not connect itself to any device OpenAI has described in public.

Why the marketing community has a stake

The application pitches silent input as a way to use devices "without having to physically manipulate the device (e.g., to type inputs or interact with buttons, etc.)" and without speaking aloud, and it is explicit that the output can go to language models and agents. That matters because OpenAI's own assistant is now an ad platform: OpenAI sells ChatGPT ads in 89 countries and said in October that more than 100,000 advertisers had tried the system since launch. On November 13, it gave its agents a wallet to buy from websites on a user's behalf. An input method that can feed "querying databases" and "composing messages" silently would sit upstream of both.

The data question is sharper. The description lists sensors for heart rate, oxygen saturation, skin temperature, hydration and facial muscle activity, and it describes inferring swallowing and muscle tension. Health readings of that kind fall under special category data in the EU, with stricter conditions for processing. Paragraph 24's reference to "miniscule" movements made "while the person is thinking about or reading words" raises a further question the application does not address: what such a device could infer about material a wearer reads, not just commands they choose to give.

Regulators have started to look hard at body-worn capture devices. On October 13, 2026, the European Data Protection Board said smart glasses makers cannot rely on wearers to meet GDPR duties. Apple restricted the listening features in watchOS 27.3 to users aged 18 and over when the first developer beta appeared on November 3, 2026. Amazon's $74.99 Bee Frontier, released on October 15, 2026, can turn products mentioned in conversation into recommendations on the Amazon homepage, which shows that ambient input from a wearable can already reach a shopping surface.

Meta has already shipped a related idea at the wrist. The Meta Neural Band, a surface electromyography wristband that comes with Meta Ray-Ban Display glasses, launched at Meta Connect on September 17, 2025, reads muscle signals instead of sound, and Meta processes the raw EMG on the device and passes only discrete click events to the glasses. OpenAI's application keeps both options open, local processing or streaming raw data to a phone or edge device, and leaves the choice to resources, battery, latency or "user preference." Separately, an academic audit published in October found that only 14 of 29 AI products let users stop their data being shared for advertising, a reminder that controls around AI inputs remain uneven.

What the document does not establish

The application reports no accuracy rates, error rates, test subjects, vocabulary sizes beyond the 44-phoneme example, latency or battery figures, and it names no product, price, launch date or manufacturing partner. It does not say whether OpenAI has built the device, whether processing would happen on the earbud or in the cloud by default, or how sensor data would be stored, used for training or protected. The claims may change before any grant.

Key facts

  • Document: US Patent Application Publication US 2026/0369273 A1, "In-Ear Silent-Speech Detection System"
  • Published: November 19, 2026 (Thursday)
  • Applicant: OpenAI OpCo, LLC, San Francisco, CA
  • Inventors: Douglas Stuart Brodie (San Jose, CA); Daniel Boothe (San Francisco, CA)
  • Application: 19/677,432, filed May 14, 2026
  • Priority: US Provisional Application 63/805,864, filed May 14, 2025, same title, claimed under 35 U.S.C. 119(e)
  • Classification: Int. Cl. G10L 15/24, G01B 7/28, G10L 15/26, G10L 15/30, H04R 1/08; CPC G10L 15/24, G01B 7/28, G10L 15/26, G10L 15/30, H04R 1/083
  • Claims: 20; independent claims 1 (motion sensing) and 10 (deformation sensing)
  • Drawings: 7 sheets; FIGS. 1, 2A, 2B, 3A, 3B, 4, 5, 6
  • Key parameters in the description: 3-axis accelerometer and 3-axis gyroscope; sample rate about 100 Hz to about 1,000 Hz; speech events below about 30, 20 or 10 dB; 44 English phonemes as a 44-entry output vector
  • Status: published application, not granted
  • Source URL: none printed on the document

Timeline

Summary

Who: OpenAI OpCo, LLC, as applicant, with inventors Douglas Stuart Brodie and Daniel Boothe. The design would affect users of earbuds and AI assistants, and companies that reach people through voice and agent interfaces.

What: A published US patent application for an in-ear device that detects words mouthed without sound by sensing how the ear canal moves (claim 1) or changes shape (claim 10) as the jaw works, then turns the signals into phonemes, words, commands or speech, with language models and agents named as possible consumers of the output. The description lists further sensors for muscle activity, heart rate, oxygen saturation, temperature and hydration.

When: Published on November 19, 2026. The provisional application was filed on May 14, 2025 and the full application on May 14, 2026.

Where: Filed with the US Patent and Trademark Office. The device would sit partly inside the ear canal and communicate with phones, computers, headsets, home devices or other wearers over Bluetooth, ultra-wideband, Wi-Fi or cellular links.

Why: The application presents silent input as a way to use devices without touching them or speaking aloud, including where speaking "would not be appropriate or desirable," and as a channel for private two-way communication. It matters to the marketing community because it describes a new input path into language models and agents at a time when OpenAI sells ads in ChatGPT and lets its agents make purchases, and because the sensors it describes collect bodily signals that privacy rules treat with particular care.