LipConfirm
Alone, a whisper can be ambiguous. Combined with your typing, it can help distinguish “live” from “love” or recover a garbled word. Typing stays primary. Fix a word with a brief cue instead of backspacing, while keeping most of your message unspoken.
All 30 claims allowed · U.S. Patent No. 12,737,540 scheduled to issue September 15, 2026.
restaurant
Generate candidates from typing first, then use lip or whisper evidence to rescore and select.
Typing only · no auxiliary evidence
Type the word to see its candidates.
Waiting for the mistyped word.
Illustrative simulation. No microphone or camera is accessed.
Typing stays primary in this simulation. Whisper and lip cues guide correction rather than provide a standalone dictation or lipreading result. A single articulation can supply both cues in a combined implementation; combined capture is not simulated here.
Even LLM-informed autocorrect can struggle with garbled input or a typo that produces another valid word. A quiet cue adds evidence of what you intended, beyond what spelling and context alone can establish.
Typing can make a faint whisper more useful than that whisper alone. Partial clues, such as syllable rhythm or a distinguishing vowel, can help select among words supported by your typing. LipConfirm uses that evidence without requiring a reliable standalone transcript. Whisper assistance needs no camera.
A slip can turn love into live or this into thus. A correctly spelled word can still be the wrong word. Both “I love my life here” and “I live my life here” are grammatical, so context may not settle your intent. The demo includes love/live. See 20 everyday examples.
Quietly articulate a word as you type it, or add a cue after you notice an error. LipConfirm can reconsider that word using the original typing. Optional silent lip assistance can use a brief, requested capture.
You may want to keep most of your message unspoken, stay quiet around other people, or simply keep using the keyboard. LipConfirm lets a brief cue assist the text you are already typing.
A very faint whisper may fail to register or may produce the wrong words at normal typing distance. Speaking louder or moving the phone toward your mouth can help, but it changes the flow of typing.
In my preliminary desktop work, combining faint whisper cues with typed input helped recover the intended word, even when the typing was garbled.
A faint cue can still carry useful information. LipConfirm is designed to use that uncertain signal together with your typing. The point at which dictation becomes unreliable varies by device and setting.
A faint whisper often retains useful clues to syllable count and rhythm, critical vowel or consonant sounds (phonemes), and sound timing. LipConfirm evaluates the available acoustic evidence together with the original typing, including garbled text. The whisper need not identify the word independently: a partial cue can help separate otherwise plausible candidates.
The capture path must preserve that weak signal. Depending on the system, speech-detection gates or audio filtering can exclude faint utterances before they reach recognition. A user-enabled typing session or requested capture can provide a path for evaluating the retained acoustic evidence with the text. A blank dictation result alone does not identify which stage failed.
Phonetic background: research on whispered speech documents vowel resonances and sound durations despite the absence of normal vocal-fold vibration. This helps explain why incomplete acoustic evidence can remain informative; it does not establish a minimum usable volume for LipConfirm.
A small keyboard slip can produce a perfectly valid word with the wrong meaning. A brief whisper or silent lip cue can add evidence of the word you intended. See what the cue contributes in selected examples below.
U I O sit side by side on a QWERTY keyboard. That makes pairs such as love/live, this/thus, if/of, and shirt/short easy to mistype. These tiny slips can be frustrating because a real word may pass a spelling check, and context does not always rescue the mistake.
| # | You meant | Typed by mistake |
|---|---|---|
| 01 | love | live |
| 02 | this | thus |
| 03 | if | of |
| 04 | shirt | short |
| 05 | pick | puck |
| 06 | sick | sock |
| 07 | hit | hot |
| 08 | tip | top |
| 09 | slip | slop |
| 10 | fill | full |
| # | You meant | Typed by mistake |
|---|---|---|
| 11 | I love my life here. | I live my life here. |
| 12 | I left it in the car. | I left it on the car. |
| 13 | I like the shirt dress. | I like the short dress. |
| 14 | We have three cats. | We have three cars. |
| 15 | Please bring the coat. | Please bring the boat. |
| 16 | The team is ready. | The tea is ready. |
| 17 | Please renew my subscription. | Please review my subscription. |
| 18 | I’ll go with the latte. | I’ll go with the latter. |
| 19 | It’s a casual relationship. | It’s a causal relationship. |
| 20 | I left it in the bag. | I left it in the bar. |
In the sentence examples, both versions are grammatical but mean different things. A language model can favor one, yet still miss your intent.
A cue need not reveal the whole word to help resolve it. Even faint whispers often carry clues to syllable count, rhythm, critical vowel or consonant sounds (phonemes), and timing. Optional lip capture can add visible closure, rounding, and movement. LipConfirm combines the available cues with the typing to help choose between plausible words; audio and lip evidence can also reinforce one another.
In “I love my life here,” a faint “uh” vowel cue can favor love over live, pronounced “liv.” The same vowel contrast helps separate thus from this. Within each pair, the consonants and syllable count match. The typing narrows the choice; the vowel adds evidence context may lack.
“I like the shirt dress” and “I like the short dress” both make sense. In common American pronunciation, the “er” sound in shirt differs from the “or” sound in short. A captured trace of that vowel can help LipConfirm reconsider the neighboring I/O key slip without transcribing the whole sentence.
If you type “The tea is ready” but articulate team, the lips close for the final “m.” A captured closure timed to that word ending can support team over tea. LipConfirm can weigh that visible cue alongside the shared typed letters; a resting mouth closure alone would not establish the word.
A captured initial “k” sound can favor coat over boat. In the other direction, boat begins with a “b” made by closing and releasing both lips, while “k” is formed farther back in the mouth. A captured lip closure and release at the word’s start can support boat among the typed candidates.
When casual is articulated with three syllables, its rhythm differs from the usual two syllables of causal. The vowel and middle consonant sounds also differ. LipConfirm can weigh those clues against the nearly identical typed letters. Pronunciation varies, so syllable count is supporting evidence rather than a fixed rule.
These words differ by neighboring I/O keys. A captured vowel cue can help favor if over of, whose vowel varies with pronunciation and emphasis. Lips alone offer less help: “f” and “v” use similar lower-lip-to-upper-teeth contact. LipConfirm can weigh the audio with the typing instead of relying on that shared mouth shape.
These walkthroughs illustrate how cues can support correction. Pronunciation and capture quality determine which evidence is available. When the combined evidence is insufficient, LipConfirm can retain the typed word or offer alternatives.
Whispered vowels can retain acoustic resonances even though a true whisper lacks normal vocal-fold vibration. Lip cues provide a different kind of evidence: “b” and “m” involve both lips, while “f” and “v” share lower-lip-to-upper-teeth contact. A visible gesture therefore supports some distinctions without uniquely identifying every sound.
Sources: Houle and Levi, acoustic differences between voiced and whispered speech; Essentials of Linguistics, consonant articulation; Cambridge Dictionary, pronunciation of “casual”. The application to these typed pairs is illustrative.
The videos below show lip-assisted correction in a desktop prototype.
In a user-enabled typing session or requested capture, the microphone can collect a brief cue without requiring a successful dictation transcript. The capture path must preserve usable faint speech. Optional camera capture can be limited to requested lip assistance, with local processing and disposal of raw data supported by the architecture.
The original typing and the acoustic or lip evidence are evaluated together. Weak cues can help distinguish between words supported by the typing. The system can rescore candidates, help recover missing candidates, or combine inputs in a learned multimodal model without first requiring a reliable standalone transcript.
The combined evidence can update a suggestion, confirm a word, or correct a valid word that was typed unintentionally. Confidence controls matter for familiar words as well as garbled input: when the evidence is insufficient, the system can retain the text or offer alternatives.
The videos below show lip-assisted correction in a desktop prototype.
Keep typing normally and add an occasional cue when you want help with a word, whether it is garbled, ambiguous, or an everyday word typed incorrectly. LipConfirm is designed to bring the combined evidence into existing keyboard, autocorrect, spell-checking, and language-model infrastructure.
Technical demonstrations and IP discussions available. Desktop prototypes cover both whisper-assisted and lip-assisted correction.
All 30 claims allowed. U.S. Patent No. 12,737,540 is scheduled to issue September 15, 2026. The allowed claim set addresses lip-assisted correction, shared multimodal operation, and a camera-free whisper-assisted pathway that generates textual proxies for an existing spell-correction engine.
Two U.S. continuation applications are pending, seeking additional protection for lip- and whisper-assisted autocorrect. One has been granted Track One prioritized examination.
PCT/US2026/035963 is pending. The USPTO, acting as the International Searching Authority, issued a written opinion with positive novelty, inventive-step, and industrial-applicability findings for all 30 claims. The international search report contains no category X or Y citations.
You may prefer typing to keep most of a message unspoken or avoid disturbing others. The whisper you are comfortable making at normal typing distance may also be too faint for reliable dictation. LipConfirm adds the evidence in your actual typed input, so the same uncertain whisper can help resolve a word without having to identify it independently. This can be useful even as dictation improves.
In my preliminary desktop work, combining faint whisper cues with typed input helped identify the intended word, including when the typing was garbled. These are early findings from selected demonstrations. I have not yet established a representative accuracy percentage across users, devices, noise levels, and normal phone typing distances. Broader evaluation needs to count both recovered words and unwanted corrections.
No. Familiar words can be mistyped into other valid words, including love/live and this/thus. “I love my life here” and “I live my life here” are both grammatical, so even a strong language model may not settle the user’s intent from context alone. A whisper or lip cue can supply additional evidence; confidence checks determine whether to change the word.
Try the progressively fainter whisper check at your usual typing distance. You may find a range where the whisper is still faintly audible to you but dictation becomes unreliable. An implementation must preserve those cues in its capture path; a blank dictation result alone does not reveal whether the failure was in detection, processing, or recognition.
You can keep most of the message unspoken and add only an occasional cue. A whisper may still be audible nearby; optional lip assistance provides a silent path. For data privacy, the architecture supports local processing and discarding raw audio or mouth-region frames. The host platform determines the implemented data handling.
Here, whisper-level audio means quiet acoustic evidence, including faint, low-amplitude or breathy speech. The aim is to use a comfortable brief cue while keeping the phone where you normally type. LipConfirm evaluates that evidence with the typed input; a complete, correct standalone transcript is not required. Usable volume depends on capture settings, microphone distance, background noise, and the model.
No. Whisper-assisted correction combines microphone input with typing and does not need a camera. Optional lip assistance can provide a silent cue or reinforce a quiet utterance.
Continuous camera use is not required. In the on-demand approach, the camera stays off until lip assistance is requested, captures a short clip of the word, then stops. Longer visual sessions remain an optional, user-enabled configuration.
Yes, in an opted-in typing session. The microphone can remain available so an occasional quiet cue accompanies a word as you type it, including a familiar word that is easy to mistype. Alternatively, it can activate only when assistance is requested. There is no requirement to whisper every word.
The architecture supports associating a new cue with the word just typed and reconsidering the correction using the retained typing evidence. Optional lip assistance can use a brief capture requested for that word.
Yes. One quiet articulation can supply audio and lip cues. The architecture combines both with typed evidence and adjusts their influence according to reliability. An unavailable or unreliable channel can contribute less or be omitted.
I can provide desktop demonstrations of whisper-assisted and lip-assisted correction. The interactive website demo is an illustrative simulation with preset examples and scores; it does not access your microphone or camera or measure live recognition accuracy.
LipConfirm is designed to support on-device correction without sending raw mouth-region frames or audio to the cloud. Raw sensor data can be discarded after feature extraction or scoring. Deployment choices depend on the host platform and model.
The design supports brief camera use and session-based or requested microphone capture. Actual battery use and latency depend on sensor settings, recognition models, and the device.
It can rescore or validate existing candidates, help generate or recover candidates, or enter a learned model together with text. A proxy-string pathway can turn speech-related clues into spelling-like inputs for an existing correction engine. The interactive demo illustrates late and early fusion as two distinct strategies.
Text-assisted correction can use partial cues, such as syllable patterns and mouth shapes, rather than independently recognizing every word from lips alone. Useful visual evidence still depends on articulation and capture quality. Optional personalization can adapt scoring to a user.
The architecture supports keyboard-based correction on phones, tablets, desktop systems, wearables, and extended-reality devices. Each platform can use the available, user-permitted combination of typed, acoustic, and visual evidence.
A high-confidence result can be inserted or confirmed automatically. When the evidence is uncertain, the system can offer alternatives, retain the current text, or wait for another cue. The user does not need to see the internal candidate scores or camera feed.
All 30 claims of the U.S. parent application have been allowed, and U.S. Patent No. 12,737,540 is scheduled to issue September 15, 2026. Two U.S. continuation applications and a PCT international application are pending. One continuation has been granted Track One prioritized examination. See Patent & IP for details.
Email: hello@lipconfirm.com
LipConfirm was conceived and developed by Andre Persidsky, a multidisciplinary engineer and inventor with experience building, patenting, and licensing technology. Andre has previously licensed technology to commercial partners including Spotify, and has led hardware/software health-technology projects in collaboration with external research institutions.