Psst… We received a tip…

You are packed into a subway train with your hands full.

You need to reply to a message.

Typing is awkward but dictating is worse. Nobody needs to hear, “I’m running ten minutes late, can you move the meeting?”

So you subtly and inaudibly mouth the words.

Your AirPods understand.

To everyone else, nothing happened. To your phone, you just spoke.

That is the promise sitting behind Q.ai, the startup Apple reportedly bought for close to $2 billion. This week we decode Q.ai’s patents which sketch out a way for wearables to detect silent speech from tiny facial skin movements, then turn that into text, audio, commands, or even translated speech.

The founder of Q.ai, Aviad Maizels, has done this before.

In 2013, he sold Apple his first company, PrimeSense, a 3D-sensing firm, for around $350 million. The technology sat inside Apple's labs for four years before it shipped as Face ID on iPhone X in 2017. Maizels joined Apple after the deal, left in 2022, founded Q.ai, and in January 2026 sold Apple his second sensor company for a reported figure somewhere between $1.5 billion and $2 billion, depending on which publication you ask. 

Q.ai's patents describe technology that enables the understanding of silent speech through analysis of microscopic movements in facial skin. There is no camera watching your mouth. What the earpiece contains is a coherent light source (aka a small infrared laser) pointed at your cheek.

When laser light hits skin, it does not bounce back cleanly. It scatters into a speckle pattern, a chaotic arrangement of light and dark produced by the microscopic surface texture of your skin. The sensor captures that pattern continuously.

As your facial muscles move, the speckle pattern shifts. The shifts are invisible to any camera in the room. Your cheek is the read surface.

When you silently form the word "call," your lips may barely part, but your cheek muscles contract, your jaw tenses, and the tissues around your mouth begin their preparation. The laser reads all of that as a sequence of changing patterns. A trained neural network matches the sequence to words.

Two details in the patent family stand out. The first is a wake-on-intent architecture. The device sits in low-power mode until it detects the early muscular signs of incoming speech, then switches to active processing. That is how an earpiece battery survives the day. The second is an ambiguity handler. Two silent words can look similar on the face. The system manages this by working with probability-weighted phoneme candidates rather than waiting for a certain answer, which makes the output usable even when the signal is messy.

The difference from lip reading matters practically. A camera needs line of sight to your mouth. This works through the cheek, in the dark, inside a stem no thicker than a pencil.

Talking to AI agents is overtaking typing.

AI-powered dictation has made transcription viable at consumer quality. ChatGPT voice mode normalized conversational AI for millions of people who had never used Siri intentionally. Apple, Google, and Meta are all racing to make their assistants speech-first.

The problem is that people won't use voice in public spaces.

Voice assistants have an enormous amount of usage in cars, at home, and alone. The social friction of speaking out loud to a device, in a meeting, on a train, in an open office, is a hard ceiling that better models don't move. Silent input removes the reason people don't use voice prompting … unless you feel like an idiot wordlessly moving your mouth in public…

Apple was not first here, and they’re paying the price.

MIT's AlterEgo project began in 2018 as a graduate research prototype. By September 2025, it had launched as a commercial company, with CEO Arnav Kapur's device using electromyography, small electrodes on the jaw and neck that detect neuromuscular signals during subvocalization. Different read surface, different method, same goal. (MIT Media Lab)

Facebook acquired CTRL-labs in 2019. After they developed a wristband that picks up signals from the brain and allows users to control a digital device without moving a finger. Meta has since evolved the technology into a wristband bundled with its Ray-Ban Display glasses, priced at $799 for the combined package. (GeekWire)

Wispr Flow is the instructive pilot. Their speech-to-text technology works even if you ‘wispr’. Founded in 2021 with the goal of building a non-invasive wearable for silent text entry, the company concluded that contemporary AI systems were insufficient and pivoted towards a voice dictation software layer. That pivot worked. As of May 2026, Wispr Flow has over 2.5 million global downloads, with aims to reach over a billion people.  (Bloomberg)

The global voice assistant category sat at roughly $7 billion in 2024, with projections to $25 billion by 2035 at the conservative end (Market Research Future) and $79 billion by 2034 at the aggressive end (Market.us). That's the market Q.ai's tech competes in.

Analyst firms cannot agree on a number for voice assistants because the category resists clean definition. Estimates for 2024 range from $2.73 billion (DataM Intelligence) to $7.35 billion (Next MSC) depending on whether you count software only, software plus hardware, or pull in customer service voice AI. The line between assistant, search, agent, and operating system keeps moving.

The broader speech and voice recognition market sits at $9.66 billion in 2025, projected to reach $23.11 billion by 2030 (MarketsandMarkets). That number captures dictation, transcription, and voice biometrics alongside assistants. Apple's silent input slots into both stacks.

The installed base is where the strategic value sits. Apple shipped an estimated 118 million AirPods units in 2025, with cumulative AirPods revenue crossing $100 billion in 2026 (Counterpoint Research). Pair that with 1.5 billion active iPhones (Business of Apps) and you have a distribution layer no startup can match. Ship silent input as a standard AirPods feature and you've rewritten Siri's input layer across an installed base larger than most countries.

Widen the lens once more. The relevant market becomes how humans give instructions to machines. Meta's $500 million to $1 billion CTRL-Labs acquisition in 2019 finally shipped as the Neural Band in September 2025, bundled with Ray-Ban Display glasses at $799 (TechCrunch). Apple Vision Pro relies on eye tracking and pinch. Synchron and Neuralink are chasing implantable brain-computer interfaces. Each bet says voice, keyboard, and touchscreen will eventually share the stage. Q.ai gives Apple a credible fourth contender.

That leaves a question worth holding onto. If voice assistants stop being audible, who else stops hearing them?

Apple confirmed the Q.ai acquisition in January 2026. The Financial Times reported the deal at close to $2 billion, which would make it Apple's second-largest acquisition after the $3 billion Beats purchase in 2014. Calcalist and Ctech placed the figure closer to $1.5 to $1.6 billion. Q.ai ran on undisclosed funding, with around 100 employees, for its entire three-year life before the deal. MacRumors

This is the second time Maizels has sold a company to Apple. In 2013, he sold PrimeSense, a 3D-sensing company, whose technology became the foundation for Face ID. TechCrunch

Facebook acquired CTRL-labs in 2019 for a figure reported between $500 million and $1 billion.(Bloomberg, CNBC). The technology has since evolved into the Meta Neural Band, bundled with Meta's Ray-Ban Display glasses. OSVR

Wispr Flow raised a $30 million Series A from Menlo Ventures in June 2025, followed by a $25 million extension from Notable Capital in November 2025, bringing total funding to $81 million and post-money valuation to $700 million (TechCrunch). Bloomberg reported in May 2026 that the company is in talks to raise a further $260 million at a $2 billion valuation.

The pattern across these deals is seven years of serious capital into a category that has not yet produced a mass consumer product. The $2bn Q.ai acquisition is the first signal that one of the incumbents is ready to ship rather than fund.

Aviad Maizels sold Apple a sensor company in 2013 and its technology took four years to appear in a product. Whether Q.ai follows a similar timeline, or ships faster because the AI layer is already ready, is the question Apple hasn't answered.

The one worth watching. Does this land in AirPods Pro or Vision Pro? AirPods has the scale. Vision Pro has the input problem. The patents describe both.

This week's core patent is US 11,915,705 B2, titled 'Facial Movements Wake Up Wearable', authored by Q (Cue) Ltd.

The broader deciphering mechanism sits in PCT WO 2023/012546 A1, titled 'Deciphering of Detected Silent Speech'.

Reply with your take. We read everything. Find us on Instagram and LinkedIn for more on the deals, patents, and market signals we're watching.