Eight PhDs walked out of a Waterloo neuroscience lab and built a chip that hears you - full vocabulary, real time - without ever calling the cloud.
Here is a thing that sounds impossible until you look closely, at which point it sounds merely difficult, which in the chip business is roughly the same thing. Applied Brain Research - ABR to almost everyone - has built a piece of silicon that recognizes full-vocabulary human speech, in real time, while consuming less power than a hearing aid. Under 35 milliwatts. The way this normally works is that your smart speaker records you, ships the audio to a data center in another province, and a rack of GPUs figures out what you said. ABR's proposition is: what if it just didn't do that? What if the whole thing happened on a chip the size of a fingernail, in the room, and your voice never left the device?
The reason this is hard is that "just run the AI locally" is the sort of sentence that hides an enormous amount of engineering. Real speech recognition wants memory, compute and power - three things that edge devices, by definition, don't have much of. The usual move is to shrink the model until it fits, at which point it gets dumb. ABR's move was different and, characteristically for this company, came out of a math paper. Instead of the transformer architecture that eats most modern AI, ABR bet on state-space models, and specifically on something called the Legendre Memory Unit, a way of compressing a stream of time - audio, sensor data, biosignals - into a compact, efficient representation. It is named after an 18th-century French mathematician whose polynomials nobody expected to end up in your earbuds.
That bet is the whole company. State-space models are fashionable in AI research now; they were not when ABR started working on them, and ABR is the first company to actually put them into silicon. There is a lesson in here about being early, which is that it is indistinguishable from being wrong right up until it isn't. ABR spent roughly a decade in the "indistinguishable from wrong" phase. In January 2026 the seed round closed, oversubscribed, and the phase ended.
Caption — The mark sits alone on a dark tile, the way a good chip sits alone on a board. Nothing extra. That is on purpose; the company that made it spends its days deleting milliwatts.
"Voice is becoming the default interface for the next wave of edge devices - but using cloud voice AI solutions is a terrible experience."KEVIN CONLEY · CEO, APPLIED BRAIN RESEARCH
ABR grew out of the University of Waterloo's Centre for Theoretical Neuroscience and Chris Eliasmith's book How to Build a Brain, which introduced the Semantic Pointer Architecture and the Neural Engineering Framework.
The team built Nengo, an open-source neural compiler and simulator used to run Spaun - one of the largest functional brain models ever built - and to interface with Intel's Loihi neuromorphic chip.
The theory hardened into the Legendre Memory Unit, then into the TSP1 Time Series Processor - real hardware, sampling to customers, purpose-built for low-power time-series AI at the edge.
The TSP1 is a general time-series engine, which is a fancy way of saying it is good at any signal that arrives as a stream over time - your voice, your heartbeat, the vibration of a motor. That makes the product list less "speech chip" and more "anything that listens."
Full-vocabulary transcription and natural, expressive speech generation - on-device, low-latency, no cloud round-trip. The privacy and battery-life trade-off simply goes away.
Real-time processing of health and bio-signals for wearables and medical devices, keeping sensitive data local by design.
Voice interfaces and sensor fusion for earbuds, hearables, glasses and home devices where power and responsiveness decide whether a feature ships.
On-device signal processing for factory sensors and machines - inference at the edge, where connectivity is unreliable and latency is expensive.
| Spec | TSP1 Time Series Processor |
|---|---|
| Speech power (ASR / TTS) | Under 35 mW |
| Full-vocabulary ASR latency | Under 35 ms |
| On-chip model capacity | Up to 10M 8-bit or 20M 4-bit parameters |
| Processing core | 32-bit RISC MCU + state-space network fabric |
| Memory | Integrated weight memory, SRAM, secure non-volatile memory |
| Interfaces | I2C, SPI, I2S, PDM, GPIO, UART · up to 4 stereo audio inputs |
| Voltage | 1.65–3.6V, integrated 0.8V core DC-DC |
| Package | 42-pin WLCSP (0.5mm) or 44-pin QFN |
| Efficiency vs. alternatives | 10–100× lower power (company figure) |
Caption — Specs are the poetry of hardware people. Read the first two rows again: cloud-grade speech, at the power of a hearing aid, answering before you finish the sentence.
A founding team drawn from Waterloo's computational neuroscience group. The unusual part isn't the credentials - it's that they turned academic theory into a shipping chip, a path most research groups never take.
Closed an oversubscribed seed round led by Two Small Fish Ventures to commercialize the TSP1 edge voice chip.
Appointed Eva Lau, co-founder of Two Small Fish Ventures, to the board of directors.
Published industry-application materials spanning AR/VR, wearables, smart home and robotics.
Announced what it describes as the world's first single-chip solution for full-vocabulary speech recognition.
Named a World Economic Forum Technology Pioneer - the only Canadian tech company selected that year.
ABR closed its seed round in January 2026, led by Two Small Fish Ventures, following the release of the TSP1 chip. The amount was undisclosed; prior aggregate funding has been reported around $3.6M USD. Capital is earmarked for scaling the chip and its state-space AI models for edge inference.
"The embedded AI market is at an inflection point. Applied Brain Research has demonstrated that sophisticated voice AI doesn't require the cloud."EVA LAU · GENERAL PARTNER, TWO SMALL FISH VENTURES
Note — Links open YouTube search results for the most current official videos and interviews.
Applied Brain Research (ABR) is a Waterloo, Ontario AI hardware company spun out of the University of Waterloo's Centre for Theoretical Neuroscience. It builds brain-inspired chips and software that run real-time AI - full-vocabulary speech recognition, text-to-speech and sensor processing - directly on edge devices at power levels measured in milliwatts. Its patented state-space models and the Legendre Memory Unit power the TSP1 Time Series Processor, which ABR calls the world's first single-chip solution for full-vocabulary speech recognition, doing the work of cloud voice AI while consuming 10 to 100 times less power.
Last updated: