# How Chips Use On-Chip Ultrasonic Transducers for Gesture Recognition

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 16 Jul 2026, 14 min.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0114.mp3

This episode explores a new frontier in human-machine interface: on-chip ultrasonic transducers that enable touchless gesture recognition. Lucas and Luna break down how these tiny MEMS devices work, from generating focused ultrasound beams to sensing Doppler shifts in reflected signals. They examine a real-world application from BMW's 2025 i7, which uses an ultrasonic sensor array to detect hand gestures for volume and navigation controls — no cameras, no privacy concerns. The hosts also discuss the chip design challenges: power consumption at 10 milliwatts, signal processing latency under 20 milliseconds, and integration with existing SoCs. A specific comparison to infrared-based systems shows why ultrasound wins in bright sunlight and through glass. The episode closes with a look at where this tech is heading — from automotive to smart home and eventually to wrist-worn wearables that replace touchscreens. #UltrasonicTransducers #GestureRecognition #MEMS #BMWi7 #TouchlessInterface #HumanMachineInterface #OnChipSensors #AutomotiveTech #SmartHome #Wearables #DopplerShift #SignalProcessing #SoCIntegration #PowerEfficiency #PrivacyTech #HardwareEngineering #FexingoBusiness #Technology Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: You know how most gesture recognition today relies on cameras — either regular RGB or infrared depth sensors like the ones in the Microsoft Kinect?
Luna: Yeah, and that works, but it has obvious privacy issues and fails in bright sunlight.
Lucas: Exactly. So there's a growing push to do gesture sensing with sound instead of light. Specifically, on-chip ultrasonic transducers — tiny MEMS devices that emit and receive ultrasound, all integrated into a single silicon die.
Luna: So instead of a camera watching you, it's more like a miniature sonar system bouncing pings off your hand.
Lucas: Right. The principle is basically the same as a bat's echolocation. A piezoelectric membrane vibrates at frequencies between 40 and 100 kilohertz — well above human hearing — and sends out a focused beam. When that beam hits a moving object, the reflected signal gets a Doppler shift. The chip measures that shift to detect motion, speed, and direction.
Luna: And all of that fits on a chip? I've seen discrete ultrasonic sensors, but they're usually separate modules.
Lucas: That's the breakthrough. Companies like TDK, Infineon, and a startup called Chirp Microsystems have been shrinking these transducers down to the point where you can integrate an array of them — say, eight by eight — directly onto a CMOS die alongside the driving electronics and signal processing logic. The whole thing is a few square millimeters.
Luna: So it's a system-on-chip for sound-based gesture recognition. What's the range and resolution like?
Lucas: Typically up to about one meter, with angular resolution of around one degree. That's enough for in-car gesturing — changing the volume or skipping a track without taking your eyes off the road. And because it's ultrasound, it works through glass and plastic, so you can embed it behind a panel or a logo.
Luna: That brings up the automotive use case. I remember BMW demonstrated something like this in their 2025 i7 — a gesture control system that didn't use any cameras.
Lucas: That's the example I was going to mention. BMW worked with a supplier called Elliptic Labs, who provide a software platform that runs on top of off-the-shelf MEMS ultrasonic sensors. In the i7, the sensor array is hidden behind the center console trim. You wave your hand left or right to skip tracks, or do a circular motion to adjust volume.
Luna: And no camera means no concern about the car recording you. That's a big selling point for privacy-conscious buyers.
Lucas: Exactly. The chip itself also consumes very little power. The transmit phase is pulsed — it only fires for a few microseconds every 50 milliseconds or so. The average power draw is around 10 milliwatts for a continuous listening mode. That's low enough to keep on even when the car is parked, so you could approach and do a gesture to unlock the door.
Luna: Is that a feature in the i7? Gesture-based keyless entry?
Lucas: Not yet, but it's in development. The latency requirement is tighter — you're aiming for under 20 milliseconds from gesture to response for it to feel natural. The current chips can achieve about 15 milliseconds with dedicated DSP hardware.
Luna: So the on-chip signal processing is key. Do they use a neural network or classical algorithms?
Lucas: Both, actually. The raw time of flight and Doppler data is handled by a simple state machine — that's the low-latency path. But for more complex gestures like a pinch or a two-finger swipe, they feed a feature vector into a tiny convolutional neural network running on the chip. The network has about 50,000 parameters — tiny by AI standards, but enough to classify 10 to 20 distinct gestures.
Luna: Fifty thousand parameters — that's practically nothing. What kind of accuracy do they get?
Lucas: In controlled lab conditions, above 98 percent. In a real car with road noise and air conditioning, it drops to around 92 percent. That's comparable to camera-based systems but without the privacy trade-off.
Luna: And it doesn't care about lighting. That's the other big advantage. Cameras struggle in direct sunlight or complete darkness.
Lucas: Right. Ultrasound is essentially immune to ambient light. It also works through minor obstructions — a thin layer of dust or a smudge on the panel doesn't degrade performance. Camera lenses fog up; ultrasonic transducers don't.
Luna: So what's holding this back from replacing cameras in phones or laptops? I've seen some Android phones with ultrasonic sensors for proximity, but not for gestures.
Lucas: The main barrier is cost and size. A single ultrasonic transducer array chip might cost three to five dollars in volume — fine for a luxury car, but hard to justify in a smartphone where every millimeter of PCB space is accounted for. Also, the beamforming requires a certain aperture; an array of 64 elements takes up about 5 by 5 millimeters. That's doable, but manufacturers have to choose it over other sensors.
Luna: And the power consumption? 10 milliwatts is a lot for a phone that's trying to last all day.
Lucas: True. But you can duty-cycle it aggressively — only turn on the transmit when you detect a hand approaching via a low-power infrared proximity sensor. That hybrid approach brings average power down to under a milliwatt.
Luna: So the chip itself is ready. It's more about system integration and willingness to replace existing sensors.
Lucas: Exactly. And I think we'll see it first in smart home devices — thermostats, smart speakers, light switches — where the privacy argument is strong and you don't want a camera pointing at your living room.
Luna: There's actually a startup called Ultraleap that's been doing this with discrete sensors for years. Have they moved to on-chip?
Lucas: They have. Ultraleap recently announced a partnership with a foundry to produce a custom ASIC that integrates 256 ultrasonic transducers on a single die, along with the beamforming and classification engine. They're targeting the XR market — think hand tracking for AR glasses without external cameras.
Luna: For AR glasses, that makes a lot of sense. You already have limited space and power budget, and cameras add weight and heat.
Lucas: And you can embed the ultrasonic array in the frame or the bridge of the glasses. The beam can reflect off the user's hand even if it's behind the glasses — not possible with a camera.
Luna: So the ultrasonic approach gives you a wider field of view, essentially.
Lucas: Exactly. The transducers can be arranged in a curved array to provide nearly 180-degree coverage. The chip handles all the beamforming and time of flight calculations on-chip, so the host processor just receives gesture events.
Luna: Is there a standard interface yet? Like, can any SoC just connect to one of these chips via I2C or SPI?
Lucas: Most of them use SPI for high-speed data transfer — the raw ultrasound data is high-bandwidth. But the chips also have a lower-speed I2C interface for configuration. The gesture output is usually a simple interrupt line plus a small packet of data specifying the gesture ID and confidence.
Luna: So it's designed to be a drop-in module. That's smart for adoption.
Lucas: Right. And the software stack is getting better. Elliptic Labs, for example, provides a library that calibrates the sensor to the specific acoustic environment — compensating for reflections from nearby surfaces. That calibration happens in milliseconds during boot.
Luna: I want to come back to the power consumption thing. Ten milliwatts continuous — is that the transmit or the total?
Lucas: That's total for a typical use case with a 10 percent duty cycle. The peak transmit power is higher, about 50 milliwatts, but because you only pulse for a few microseconds, the average stays low. The receive electronics are always on, but they draw only a few hundred microwatts.
Luna: And the DSP? The neural network inference?
Lucas: That adds maybe another two milliwatts when active. But you can gate it — only run inference when the simple state machine detects a potential gesture. So the average power for a smart home device that wakes on gesture might be under 5 milliwatts.
Luna: That's low enough for battery-powered devices. A typical smart home sensor runs on a coin cell for a year.
Lucas: Exactly. And that's where I think the volume will come from — not cars, but millions of light switches and thermostats that respond to a wave of your hand.
Luna: Funny to think that the same basic physics as a submarine's sonar is ending up in a light switch.
Lucas: But that's the story of MEMS — miniaturizing macroscopic sensors onto a chip. We've done it for accelerometers, gyroscopes, microphones. Ultrasonic transducers are just the next step.
Luna: Speaking of next steps — are there any research directions that could make these chips even more capable?
Lucas: Two big ones. First, 3D imaging. Current arrays give you 2D position and velocity. Researchers at Stanford have demonstrated a 32-by-32 element array that can reconstruct a 3D point cloud of a hand — similar to a depth camera but using ultrasound. That could enable much more expressive gestures.
Luna: But that requires a lot more compute. 1024 channels of ultrasound data.
Lucas: Right. So the second direction is integrating the beamforming directly in the analog domain, using phase shifters on each element. That reduces the digital processing load dramatically. A startup called Vesper Technologies is working on that — they have a prototype that does 3D gesture recognition with only 64 elements and a few milliwatts of DSP power.
Luna: So we could see 3D gesture recognition in a smartwatch within a few years.
Lucas: That's the goal. Imagine flicking your wrist to dismiss a notification without touching the screen.
Luna: Which brings up another point: haptic feedback. Could these same transducers be used to generate focused ultrasound for mid-air haptics?
Lucas: Absolutely. The same MEMS transducers can be driven at higher power to create acoustic radiation pressure — you can actually feel a 'virtual button' in mid-air. That's a different chip design because you need more power handling, but several groups are working on a combined transmit-receive array that does both gesture sensing and haptic feedback.
Luna: So one chip could detect your gesture and then give you tactile confirmation — all without touching anything.
Lucas: Right. And that would be a game-changer for public touchscreens — ATMs, kiosks, elevator panels. You could interact without physically contacting a surface, reducing germ spread.
Luna: That feels like a clear post-pandemic use case that hasn't really taken off yet.
Lucas: The pandemic accelerated touchless interfaces, but camera-based systems raised privacy concerns. Ultrasound solves both problems.
Luna: Let's talk about the chip itself for a second. What's the manufacturing process? Is it standard CMOS?
Lucas: It's a mems on cmos process. The transducer membranes are typically made of aluminum nitride or PZT — piezoelectric materials. You fabricate the CMOS electronics first, then deposit and pattern the MEMS layers on top. That's a mature process at fabs like STMicroelectronics and TSMC.
Luna: So the cost per chip is coming down as the process matures.
Lucas: Exactly. Five years ago, a single-chip ultrasonic array cost around 15 dollars. Now it's in the three-to-five dollar range. At scale, it could drop below two dollars.
Luna: That would make it competitive with infrared gesture sensors, which are around a dollar fifty in volume.
Lucas: But infrared has the sunlight problem and the privacy problem. For many applications, the extra dollar is worth it.
Luna: Alright, so let's wrap up with what you think the adoption timeline looks like.
Lucas: I think we'll see ultrasonic gesture control become standard in premium cars within two years — BMW, Mercedes, maybe Tesla. Smart home devices will follow in three to five years. And then AR glasses and wearables in five to seven.
Luna: And you're betting that privacy will be the driving force, not technical superiority.
Lucas: Partly. But also reliability. Ultrasound works in conditions where cameras fail. And once the cost drops, there's really no reason not to use it.
Luna: One quick thing before we go — and this ties into what you said about reliability and long-term adoption. This show is only possible because a small group of listeners chip in each month through buy me a coffee dot com slash fexingo. It keeps the whole thing ad-free and lets us dig into topics like this without chasing clicks.
Lucas: Yeah, it's a real help. And if this episode gave you a new angle on gesture tech, that's exactly what we're here for.
Luna: Alright, so back to the future: do you think ultrasonic will eventually replace touchscreens entirely?
Lucas: Not entirely. But for a lot of interactions — especially in cars, homes, and public spaces — it's the better fit. The chip is ready. Now it's about the world catching up.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.