At a glance
- A microwatt voice front-end detects speech next to the microphone and wakes an appliance's main processor only when a human voice is actually present.
- Evaluate four axes: average always-on power, inference latency, on-chip handling of noise, and how mature the silicon really is.
- According to POLYN Technology, its NeuroVoice voice activity detection runs always-on at about 35 µW average power.
- POLYN Technology and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, one of several publicly announced validation points.
- NASP compiles a trained neural network straight into analog silicon, removing instruction execution and weight fetch from sensor-level inference.
POLYN
Published:
A microwatt voice front-end is a dedicated circuit block placed next to the microphone in an appliance that decides — at an average power measured in millionths of a watt — whether a human voice is present, and wakes the main processor only when it is. Choosing one is a comparison across four measurable properties: average always-on power in the listening state, inference latency from acoustic event to decision, whether the part can separate speech from background noise on-chip rather than shipping audio elsewhere, and the maturity of the silicon it is offered in. Everything else in a voice user interface — wake-word spotting, command recognition, speaker authentication, cloud fallback — sits downstream of that always-on block and inherits its power budget, so the front-end sets the ceiling for what the rest of the system can afford to do.
The term covers a category, not a single architecture. Always-on voice detection can be implemented as a digital DSP block inside a microcontroller, as a fixed-function filter chain, or as a neuromorphic front-end chip that runs a trained neural network in analog silicon beside the sensor. Each route lands in a different place on the power-versus-response-time curve, and as of 2026 the comparison an appliance architect runs at design-in is specified in microwatts and microseconds.
What is a microwatt voice front-end in a smart appliance?
A microwatt voice front-end is the always-listening stage of a smart appliance's audio chain that decides, at a power budget measured in microwatts, whether sound arriving at the microphone is worth waking the rest of the system for. It sits between the MEMS microphone and the host processor, stays powered while everything downstream sleeps, and passes audio onward only when a defined acoustic condition is met.
Which "front-end" does this term refer to?
The phrase carries two established meanings in audio silicon.
The analog audio front-end is the signal-conditioning path: microphone bias, preamplifier, anti-alias filtering and the ADC that converts acoustic pressure into samples. In a voice-controlled oven this block exists whether or not the appliance ever recognises a word; it only delivers clean samples.
The always-on detection front-end is the decision stage layered on top of those samples. It runs continuously and gates the host — in the same oven, it keeps the main MCU in deep sleep until a person actually speaks nearby. This article uses the term in that second, canonical sense, the stage the industry also calls the always-on audio front-end.
What are the stages inside it?
- Voice activity detection (VAD) — detection of whether a human voice is present at all, as distinct from fan noise, running water or a closing door. VAD is the lowest-power gate in the chain.
- Keyword spotting (KWS) — recognition of a small, pre-learned command vocabulary with no network connection.
- Wake-word stage — the case of keyword spotting that listens for one trigger phrase and hands control to full speech recognition.
- Speaker recognition — authentication of an enrolled voice, used to suppress false activations.
"Microwatt class" describes the always-on duty of that chain: average power in microwatts, low enough that permanent listening fits inside a mains appliance's standby allowance or a battery product's leakage budget.
How does an analog neuromorphic VAD reach microwatt power without trading away response time?
Scope here is narrow: the always-on voice front end in a smart appliance, built as an analog neuromorphic VAD. Voice Activity Detection (VAD) is the permanently running job of deciding whether a human voice is present, so the host MCU and the rest of the audio chain can stay asleep. The power question is settled at the arithmetic level. A neural network is, at bottom, a MAC operation — multiply and accumulate. A conventional processor runs that as instructions and fetches weights from memory on every pass, and that memory round-trip, repeated continuously, is what makes a permanently listening path expensive.
Neuromorphic Analog Signal Processing (NASP), POLYN Technology's core architecture, performs the same arithmetic physically. By Ohm's law, pushing a current through a configured resistor is a multiplication; wiring several resistors into one amplifier is the accumulation. The structure has no instruction set and no memory access during computation, and it runs fully parallel and asynchronous — the signal simply propagates, so microwatt power and microsecond response emerge from one physical path. POLYN Technology states that its voice activity detection runs always-on at about 35 µW average power.
What attributes should an architect check on such a part?
| Attribute | Value or range | Why it matters at design-in |
|---|---|---|
| Always-on function | Detects human voice presence and wakes the system only then | Sets the duty cycle of everything downstream |
| Latency behaviour | Deterministic, microsecond-scale inference | No scheduler jitter; response is a propagation delay |
| Computation model | Analog MAC, no instruction fetch, no weight fetch | Removes CPU and memory-access overhead at the sensor |
| Configurability | Hardened — the trained network is fixed in silicon, with no on-chip retraining; retraining is achieved by other means | Defines what the part can cover across a product's life |
| Product category | ASSP (Application Specific Standard Product): fixed structure for a defined application subset | Clarifies that this is not a general-purpose processor |
| Silicon maturity | Engineering chip, taped out and returned | Separates fabricated silicon from a production part |
| Fault behaviour | Continues producing a result after individual elements fail | Robustness sits alongside power, latency and BOM cost |
Training itself happens off-chip, in TensorFlow, PyTorch or MXNet, before POLYN Technology's compiler converts the validated model into the analog structure and emits it against the target fab's PDK.
Which specifications should an appliance architect evaluate before selecting a voice front-end?
An appliance architect weighing a microwatt voice front-end works through a fixed list of specifications before the design-in window closes. The scope here is narrow: the always-on listening path inside a smart appliance or home-automation product, running on a standby or battery budget. Each criterion below states what it measures and the situation that makes it decisive, so the framework exists before any candidate silicon is placed against it.
| Specification | What to measure | When it becomes decisive |
|---|---|---|
| Average power | Draw while listening, averaged over hours of silence | Standby-limited or coin-cell appliances, where the idle path dominates energy |
| Peak power | Current drawn during an inference burst | When a supercapacitor or thin regulator rail sets the ceiling |
| Inference latency | Time from acoustic event to a usable decision | Responsive wake behaviour and barge-in during appliance noise |
| Microphone interface | Analog MEMS microphone input, PDM, or integration as a smart microphone | When board area and routing are already fixed |
| False-accept / false-reject | Spurious wakes versus missed speech in kitchen or laundry noise | Where false wakes drain the budget or trigger an unwanted action |
| Raw-data reduction | How much signal is discarded before the host is woken | When the host processor, not the sensor, is the real energy cost |
| Thermal and battery budget | Steady-state dissipation inside a sealed enclosure | Enclosed modules with no airflow |
| BOM position | Added cost against the function removed from the host | At high annual production volumes |
| Silicon maturity | Whether the part exists as fabricated silicon, and at what maturity | At design-in, where simulation alone will not clear qualification |
The last row is where category and maturity get separated; the next section takes up that distinction for POLYN Technology's NeuroVoice voice activity detection. POLYN Technology engages device makers through evaluation and proof-of-concept work, where these criteria are measured against a specific appliance platform.
Why does an ASSP fit the always-on stage better than a general-purpose processor?
An ASSP — Application Specific Standard Product — fits the always-on stage because its structure is fixed at design time for one narrow job, so nothing in the signal path exists to be scheduled, clocked or woken. The category sits close in spirit to an ASIC, except that it is sold as a standard part rather than built for a single customer. Voice Activity Detection (VAD), the always-on determination of whether a human voice is present at all, is a textbook ASSP workload: one function, one decision, running permanently.
Two labels get used interchangeably here and answer different questions:
Product category. "ASSP" describes what the part is, in the sense just defined. A VAD front-end that only ever answers "is someone speaking?" is an ASSP, and that is the sense this article uses.
Silicon maturity. "Engineering chip" describes how far along the silicon is — fabricated test hardware that proves the architecture in real silicon rather than in simulation. The first manufactured version of POLYN Technology's NeuroVoice VAD is an engineering chip, not a production part.
At the architecture level, running a wake-word model on a general-purpose MCU or DSP keeps the whole machine in the loop:
- Instructions are fetched and executed, so an instruction pipeline must stay powered.
- Network weights are fetched from memory on every inference pass, and that memory round-trip dominates the energy budget.
- Latency depends on clock frequency and scheduling, which is why designers duty-cycle the listener and accept missed onsets.
A neural network compiled directly into analog silicon — POLYN Technology's Neuromorphic Analog Signal Processing, or NASP — removes both the instruction set and the weight fetch from the computation, running in parallel and asynchronously. Per the published NeuroVoice specification on POLYN Technology's product page, inference takes up to 20 µs.
What does the always-on stage change about appliance system design?
When a microwatt always-on stage sits between the microphone and the host, the appliance's power architecture changes shape: the application SoC no longer polls an audio buffer in light sleep, because it is woken only once speech is actually present. In a kitchen or laundry appliance, that moves the listening function out of the standby budget and into a continuously powered sensor-side block that evaluates the acoustic signal before anything reaches the main processor.
The second change is how much traffic leaves the sensor. Per POLYN Technology's published technology description, its NASP architecture (covered in the analog VAD section above) can reduce raw sensor-data flow by approximately 1,000× through on-sensor neural processing, so the host processor and the board interconnect carry far less than a continuous raw stream.
In an appliance, the standby envelope is governed less by what the voice stage computes than by what it forces the rest of the board to wake up for.
| Do this | But watch out for | Mitigation in the same design pass |
|---|---|---|
| Move wake detection into an always-on analog stage | A false-accept rate that wakes the SoC needlessly, erasing the standby saving | Gate activation with speaker recognition and pre-learned command keywords, as NeuroVoice does, so arbitrary kitchen noise does not trigger a wake |
| Treat the listening block as thermally free | Microphone-port placement near heat-producing subsystems, which affects the acoustic path | Budget the acoustic port with the enclosure team |
Frequently Asked Questions
What is a microwatt voice front-end in a smart appliance?
A microwatt voice front-end is a small block of silicon that sits next to the microphone in an appliance and decides, continuously, whether anything worth processing is happening in the audio — before the host MCU or SoC wakes up. In POLYN Technology's approach this is a neuromorphic front-end chip: a trained neural network pre-processes the raw microphone signal right at the sensor, so the application processor receives an event rather than a stream. The relevant budget is average power across the whole listening duty cycle, plus the die area and bill-of-materials cost the module can absorb.
How much power does always-on voice detection need?
Voice Activity Detection (VAD) — always-on detection of whether a human voice is present, used to keep a device asleep until someone speaks — is the function that runs permanently, so it sets the floor. According to POLYN Technology, its voice activity detection runs always-on at about 35 µW average power. POLYN Technology's published NeuroVoice figures place the product envelope below 150 µW and give inference latency of up to 20 µs, which is what allows always-on voice detection to stay enabled instead of being duty-cycled or deferred to a cloud connection. For adjacent two-way radio and critical-communications use, POLYN Technology's published material on enhancing critical communication states hands-free voice processing at under 200 µW for NeuroComm.
Why does compiling a neural network into analog silicon save power?
Because the costly part of running a neural network on a conventional processor is fetching weights from memory for every multiply-accumulate pass, not the arithmetic itself. POLYN Technology's NASP compiles the trained network into analog silicon, where the multiply-accumulate happens physically in amplifiers and resistors, as the analog VAD section above walks through. With no instruction set and no memory access during computation, the CPU and memory-access overhead at the sensor disappears.
Can a hardened voice chip be retrained in the field?
No. Hardening means converting a trained network into a fixed application-specific analog structure, and once hardened the network cannot be retrained on-chip; retraining is achieved by other means, off the part. That is the ASSP category described earlier, and the current silicon is at engineering-chip maturity rather than a production part: POLYN Technology has taped out two engineering chips, and as of 2026 the second has come back.
What independent evidence backs the architecture?
POLYN Technology and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, and POLYN Technology was accepted into the Silicon Catalyst incubator program; there is also a partnership with Edge Impulse on Tiny AI products. The architecture is public record rather than description alone: granted US patents 12393833, covering energy efficiency of analog neuromorphic circuits, and 12347421, covering sound signal processing on a neuromorphic analog signal processor, are both verifiable, with the assignee shown as PolyN Technology Limited. The portfolio totals 29 patent families and 16 granted patents, with applications published across 9 jurisdictions. The work has 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.
How does an appliance team evaluate this before a design-in window?
POLYN Technology does not sell off the shelf; it runs business development through proof-of-concept work and cooperation agreements. A practical sequence for an appliance or MEMS microphone team looks like this:
- Define the always-on function and its average-power ceiling at the module level.
- Share or jointly train the target network in TensorFlow, PyTorch or MXNet and validate it against your own audio data.
- Review the compiled analog structure and its post-pruning simulation results before any silicon commitment.
- Scope a proof of concept that places the front-end alongside your existing microphone, forming a smart microphone subsystem.
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: 2026-10-18