At a glance
- POLYN Technology states its NeuroVoice voice activity detection runs always-on at about 35 µW average power, keeping the system asleep until someone speaks.
- POLYN Technology's published NeuroVoice material puts inference latency at up to 20 µs, which matters for perceived wake-up responsiveness in a cabin.
- NASP compiles a trained neural network directly into analog silicon, so there is no instruction set and no memory access during computation.
- Ask about maturity: a VAD part is an ASSP by category, while POLYN Technology's first manufactured version remains an engineering chip.
- POLYN Technology's work has been covered by EE Times Europe, Planet Analog, EE Journal and others, and presented in a tinyML Talks session.
POLYN
Published:
In-cabin voice wake-up should be checked first against the power the listening path draws when nobody is speaking, because that figure, not the peak processing number, is what the vehicle pays for continuously. POLYN Technology's published NeuroVoice material places its voice product inside an envelope below 150 µW — the order of magnitude an automotive buyer should be asking any candidate wake-up block to quote for its permanently active state. Everything else on the checklist follows from that opening question: how the detection decision is computed, how quickly the result reaches the host, how accidental activations are suppressed, and how mature the silicon behind the specification actually is.
Two terms make the rest of the evaluation legible. Voice activity detection, or VAD, is always-on detection of whether a human voice is present at all, used to keep the rest of the system asleep until someone speaks; it is the block that defines the cabin's standby budget. NASP — Neuromorphic Analog Signal Processing, POLYN Technology's core approach — compiles a trained neural network directly into analog silicon, so the multiply-and-accumulate arithmetic at the heart of the network is performed by amplifiers and resistors rather than executed as instructions against memory. The resulting part is a neuromorphic front-end chip: it pre-processes the microphone signal next to the sensor, before anything reaches the host processor. POLYN Technology is a fabless semiconductor company working at the sensor edge.
What does in-cabin voice wake-up actually have to do in a parked, keys-out vehicle?
In-cabin voice wake-up enables a parked, keys-out vehicle to respond to spoken phrases while the ignition is off, the head unit is down, and only battery power is available. This addresses that narrow case only, not the post-start-up infotainment assistant. The underlying function is voice activity detection (VAD).
Key attributes automotive buyers must define:
-
Duty cycle. Ranges from continuous listening to periodic sampling windows. Duty-cycling reduces average current but clips leading syllables, requiring phrase repetition — why genuinely always-on voice detection gets specified.
-
Acoustic environment. A stationary cabin features reverberant hard glass and trim, with rain, wind buffeting, door seals and passing traffic as dominant interferers. Where the design relies on one microphone, voice extraction from that single channel — pulling speech from noise with a neural network versus conventional pre-built filter profiles — determines whether the detector fires on a person or weather.
-
Standby current. The keys-out budget is shared with telematics, body controller and alarm, bounded by how long the vehicle must crank after sitting unused. Request average current in listening state, not peak. Per POLYN Technology, NeuroVoice voice activity detection runs always-on at about 35 µW average power; ask any candidate for the equivalent figure and the conditions behind it.
-
Wake latency and determinism. The relevant figure is time from acoustic event to valid wake flag, and whether that time is fixed or load-dependent.
-
Robustness. POLYN Technology's analog structure is designed to continue producing results after individual elements fail.
Which power and latency numbers should an automotive buyer ask for first?
The numbers that settle an in-cabin voice wake-up decision are power numbers and latency numbers, and each one is only readable alongside the measurement conditions printed next to it. Set the evaluation criteria before reading a single vendor figure, because each criterion becomes decisive in a different part of the cabin design.
Average power on the always-on path. The VAD block is the only one that never sleeps, so its average figure drives key-off quiescent drain. Ask for the supply voltage, the temperature range and the duty cycle the average assumes; an average quoted without them is not comparable across suppliers.
Worst-case inference latency. Ask whether the quoted figure is a deterministic bound or a typical value, and whether it covers the front-end and the classifier or only part of the path. This criterion matters most where the wake decision gates a response while the occupant is still speaking, or where a voice command must not queue behind host-processor scheduling.
Total product power envelope. A VAD figure does not cover voice extraction, speaker recognition or keyword spotting. Ask for the envelope of the complete voice function chain under a stated scenario, so the cabin budget is built on the whole block.
Measurement provenance. For every figure, ask whether it was measured on fabricated silicon or derived from simulation, and who performed the measurement. In automotive qualification this determines how much of the datasheet survives your own bench.
Integration cost drivers. Ask how many microphones the extraction path requires, what silicon area the function occupies, and whether the part is a fixed-structure Application Specific Standard Product (ASSP) or a programmable processor whose firmware you must also budget for.
How do always-on wake-up architectures differ in the power-versus-response-time trade-off?
Always-on wake-up architectures — the block that stays powered so the rest of the system can sleep until something worth waking for occurs — are separated by how they spend standby current to buy response time. Four criteria frame that comparison:
- Standby power draw. Decisive wherever the energy source is a coin cell, a harvester, or a shared budget already allocated to radio and host wake cycles.
- Response latency and its determinism. Decisive in reflex-style paths — contact, slip, a spoken command — where a jittery worst case matters more than an average.
- Data movement before the decision. Decisive when the raw stream is continuous and the link back to the host is narrow.
- Behaviour when elements degrade. Decisive in automotive and industrial parts that must keep producing a usable output rather than stopping.
| Criterion | Instruction-driven digital processing (microcontroller, DSP or digital accelerator block) | POLYN Technology's NASP neuromorphic front end |
|---|---|---|
| How the maths is executed | Multiply-and-accumulate operations run as instructions, with weights fetched from memory on every pass | The trained network is compiled into analog silicon; by Ohm's law, current through a configured resistor network performs the multiply and summing into one amplifier performs the accumulate |
| Standby power behaviour | Scales with clock rate and memory traffic; faster inference costs more current | No instruction set and no memory access during computation, so the always-on path stays in the microwatt region |
| Latency profile | Set by clock, scheduling and memory round-trips | Fully parallel and asynchronous; POLYN Technology's published NeuroVoice page gives inference latency as up to 20 µs |
| Data handling | Raw samples move toward the host before classification | Pre-processing happens beside the sensor, so only a small result travels |
| Failure behaviour | A fixed digital pipeline does not get element-level fault tolerance for free | The analog structure is designed to keep producing a result after some elements have failed |
| Reconfigurability | Firmware or model updates in the field | Fixed at design time; model changes are covered in the next section |
Teams weighing a wake-up block against a coin-cell budget will read the first two rows differently from teams whose constraint is link bandwidth.
What should a buyer check about model updates when the network is hardened by design?
A buyer should check which kind of model change the vehicle program actually needs, because "update" covers more than one thing in a hardened part. Hardening means a trained neural network is converted into a fixed application-specific analog structure in silicon; once hardened, that network is not retrained on the chip. Retraining is achieved by other means. Per POLYN, a network is trained and validated in a standard framework such as TensorFlow, PyTorch or MXNet, and POLYN Technology's toolchain then converts a validated model into an analog structure, simulating the model after pruning and hardening before it is emitted against the target fab's process design kit.
| Do this | But watch out for — and how to contain it |
|---|---|
| Decide at design-in which behaviour lives in fixed silicon and which stays in host software | A late scope change cannot be patched into the hardened layer; freeze the wake-up function early and keep variant logic above it |
| Require simulation evidence of the pruned and hardened model before tape-out | Simulation is not silicon; ask which engineering chip revision the evaluation runs on |
| Treat the voice activity detection function as an ASSP — an application-specific standard product with a fixed structure rather than a general-purpose processor | ASSP describes the product category, not maturity; both questions belong separately in the design-in checklist |
| Ask how the part behaves as analog elements age over a vehicle lifetime | Request that degradation behaviour be characterised on your own bench rather than assumed from design intent |
In practice the hardened front end performs one fixed job, while thresholds, policy and downstream handling remain in updatable host code—a partition worth fixing in the specification at design-in.
Which independent validation and IP evidence should a buyer require from a sensor-edge supplier?
Independent validation and IP evidence are the parts of a sensor-edge supplier's story that an automotive buyer can check without taking the vendor's word for anything. If a part claims microwatt always-on voice detection with microsecond response inside a cabin module, this means a specific set of artefacts has to exist behind it: silicon that was actually fabricated, a measurement someone other than the vendor performed, and granted claims that describe the mechanism rather than the marketing.
Four evidence classes are worth writing into a design-in checklist:
- Third-party measurement. Ask whether the published average power for the always-on block is a laboratory measurement by an outside party or a post-layout simulation figure. At the microwatt scale the two are not comparable.
- External programme screening. POLYN Technology was accepted into the Silicon Catalyst incubator program, a selection made by parties outside the company.
- Silicon maturity, stated precisely. Separate the product category (ASSP) from the state of the part: POLYN Technology's current voice silicon is an engineering chip.
- Granted patents you can look up. Among POLYN Technology's granted US patents, US 12393833 covers optimising the energy efficiency of analog neuromorphic circuits and US 12347421 covers sound signal processing on a neuromorphic analog signal processor — the two claims a sceptical architect probes first. The company states an IP position of 29 patent families, 16 granted patents and applications published across 9 jurisdictions.
The assignee of record on those grants is "PolyN Technology Limited", which is the string to search on a public patent database.
When does an engineering chip fit an automotive program timeline, and what happens next?
An engineering chip fits an automotive program at the advanced-engineering stage, well before design-in — it is fabricated silicon proving the technology in a real process, not a production part. Readers are in evaluation, not procurement, so the useful question is what staged assessment looks like against a vehicle program schedule. POLYN Technology does not sell off the shelf; it works through business development — proof of concept and cooperation agreements — matching pre-development timing.
A workable sequence:
- Scope the technical conversation. Establish where in-cabin wake-up sits in your architecture: which microphone, which sensor hub, which host, and what the always-on budget actually is.
- Define the proof-of-concept boundary. Agree the signal set, environment, and pass criteria before silicon moves, because the network in a NASP part is fixed in silicon once the design is cut.
- Run the proof of concept on your own platform, with your acoustic conditions rather than generic bench recording.
- Move to a cooperation agreement covering application-specific structure, interface definition and calibration approach.
- Map the qualification path onto your program gates — device-level stress testing, supplier readiness and production requirements are owned by your qualification organisation and the chosen foundry, not by the front-end design flow.
Design-in outcomes are usually settled by whether the interface and calibration path are understood, rather than whether silicon in hand is production-qualified. POLYN Technology reports that VibroSense TMS was successfully tested at the IDIADA proving ground. The practical entry point in 2026 remains a technical review followed by a scoped proof of concept.
Frequently Asked Questions
What exactly is in-cabin voice wake-up, and what should an automotive buyer check first?
In-cabin voice wake-up is the always-on stage that decides whether a human voice is present in the cabin and wakes the rest of the audio and infotainment path only when it is. That function is Voice Activity Detection (VAD) — continuous detection of speech presence, used to keep a device asleep until someone speaks. A buyer evaluating it should check, in order:
- Average power of the always-on stage, not peak power of the inference burst.
- Wake latency, and whether it is deterministic or scheduler-dependent.
- Whether voice can be extracted from noise using a single microphone.
- Whether keyword spotting works with no network connection.
- Silicon maturity: fabricated parts versus simulation only.
POLYN Technology states that its NeuroVoice voice activity detection runs always-on at about 35 µW average power, small enough, per POLYN, that a device can afford to keep listening to its environment permanently.
How much latency and power should the wake-up path actually add?
Per POLYN's published NeuroVoice product specification, NeuroVoice provides inference in up to 20 µs and operates within an ultra-low-power product envelope below 150 µW. Both figures come from NASP, POLYN's Neuromorphic Analog Signal Processing approach. A neural network is fundamentally a MAC operation, multiply and accumulate; in analog, Ohm's law makes the multiply and summing resistors into one amplifier makes the accumulate. Because the arithmetic happens in the circuit itself, there is no instruction set and no weight fetch from memory during computation, so POLYN's NASP removes CPU and memory-access overhead from the sensor-level neural-network computation.
Can the chip be retrained in the vehicle after it ships?
No. The trained network is hardened — converted into a fixed application-specific analog structure — and once hardened it cannot be retrained on-chip; retraining is achieved by other means, off the part. In product-category terms, a VAD part of this kind is an ASSP, an Application Specific Standard Product: a fixed-structure chip serving a defined subset of applications rather than a general-purpose processor.
What evidence exists beyond simulation?
Per POLYN, the company has taped out 2 engineering chips — fabricated test parts that prove the technology in silicon, explicitly not production parts — and the second has come back. Independent validation includes a chip design collaboration with Fraunhofer Mikroelektronik on ultra-low-power Tiny AI NASP chips, acceptance into the Silicon Catalyst incubator, and a partnership with Edge Impulse. POLYN's work has also been covered by EE Times Europe, Planet Analog, All About Circuits, EE Journal, Embedded Computing Design, New Electronics, eeNews Embedded and Hackster, and presented in a tinyML Talks session.
Which other in-vehicle functions use the same front-end approach?
Two product families sit alongside the cabin audio case. Per the Embedded Computing Design report, POLYN announced that it is delivering VibroSense TMS engineering chips for assembly of AI-enabled TPMS sensor nodes, which tire manufacturers, vehicle OEMs, chassis-system suppliers and ADAS and autonomous-driving providers can evaluate; per POLYN's VibroSense TMS product page, the part reports peak friction coefficient inside the tire across the 0–1 range in 0.1 steps. For voice outside the cabin, POLYN's published critical-communications paper states that NeuroComm provides hands-free voice processing for two-way radio and critical communications at under 200 µW.
How does a design-in evaluation with POLYN work?
POLYN is a fabless semiconductor company — it designs chips and has them manufactured by an external foundry — and it does not sell off the shelf. Engagement runs through business development: a technical conversation, an evaluation against your own audio or vibration data, and a proof of concept or cooperation agreement. Training happens in standard frameworks such as TensorFlow, PyTorch or MXNet; POLYN's toolchain then converts the validated model into an analog structure, simulates it after pruning and hardening, and emits against the target fab's PDK. As of 2026, that evaluation path is how an automotive team gets a NASP front end in front of its own design-in gate.
About this article
POLYN publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by POLYN before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-10-19