The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Yes—speech recognition and natural-language processing can run locally on a microcontroller, but not every task needs to run on the same processor. In an EE Times interview published November 14, 2025, Infineon’s Omar Cruz described PSoC Edge as a two-stage system: a low-power path listens for acoustic activity and wake words, then a higher-performance processor and neural-network accelerator can handle more demanding language tasks. The interview presents this as a way to reduce network dependence, latency, and audio sent to the cloud; its power figures are vendor-reported ranges, not independent benchmarks.
How PSoC Edge splits always-on listening from language processing
The design separates a relatively simple, continuous job—acoustic activity, wake-word, and keyword detection—from heavier natural-language processing (NLP). The low-power domain can stay active while the higher-performance domain sleeps. When it detects a wake event, the system can enable a Cortex-M55 path with Helium DSP capability and an Ethos-U55 neural-processing unit for the more demanding work.
This division is intended to avoid keeping the full compute path active just to listen for a short trigger phrase. It also means that “runs speech locally” does not imply every part of an assistant uses the same processor or draws the same power: the listening stage and the follow-on NLP stage have different workloads.
Infineon says all four PSoC Edge variants include its NN Light accelerator. The E83 and E84 add more advanced neural-network acceleration. E82 and E84 add 2.5D graphics, and E84 adds extra SRAM. The interview does not provide a complete specification comparison for every variant, so these distinctions should not be treated as a substitute for checking the selected device’s datasheet.
#1 Best Overall
- Development Board N76E003AT20 Development Board System Board Core Board Minimum System Module DIY Electronic
| Variant or group | Features stated in the interview |
|---|---|
| All four PSoC Edge variants | Infineon NN Light accelerator |
| E83 and E84 | Additional, more advanced neural-network acceleration |
| E82 and E84 | 2.5D graphics |
| E84 | Extra SRAM, in addition to the features listed for its group |
What the power claims do—and do not—tell you
Cruz described always-on wake-word detection and keyword spotting as using “single digit milliwatts,” with the exact figure dependent on the use case. He characterized NLP as operating in the milliwatt range, potentially reaching hundreds of milliwatts for more demanding stages. These are high-level vendor statements from the interview, not published independent measurements.
| Workload | Infineon’s interview description | What is not specified |
|---|---|---|
| Always-on acoustic activity, wake-word, and keyword detection | Single-digit milliwatts, depending on use case | Exact model, microphone conditions, clock settings, measurement method, and reproducible test setup |
| More demanding NLP | Milliwatt range, potentially hundreds of milliwatts for heavier stages | Specific workload, model, latency, clock settings, and measured power for a defined configuration |
Those ranges help explain the intended architecture, but they are not enough to estimate battery life or compare devices. For an actual product decision, measure the complete listening and response cycle using the intended microphone, model, clock settings, and duty cycle. Compare alternatives under the same conditions rather than treating a broad vendor range as a like-for-like result.
Why run speech inference on the device?
- Less dependence on a network round trip: A local response does not need to wait for audio to travel to a cloud service and back. The episode and its companion EE Times YouTube listing describe low latency or “minimal latency,” but provide no measured latency benchmark.
- Audio can remain local: Infineon presents endpoint processing as a way to avoid sending audio to the cloud, describing the benefit as “zero data egress.” That is a platform use case, not a guarantee about every product: device logs, telemetry, app behavior, and other integrations still determine what data leaves a finished system.
- Basic functions can work offline: A device with the relevant local models can continue those functions without network connectivity. The interview does not establish that every assistant feature or every model works offline.
The companion EE Times YouTube listing, published January 8, 2026, also frames the discussion around on-device NLP, power, latency, and privacy. These are the stated benefits of local inference; the episode does not include comparative tests against a cloud assistant.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhat model size and applications does Infineon describe?
Cruz said that an edge language model with more than 25 million parameters—Infineon, 2025—can run on PSoC Edge. The interview does not identify a particular model, quantization, accuracy result, or response-time result for that statement. Parameter count alone therefore does not show how useful a model will be on a particular device or workload.
Examples discussed include an offline assistant in a smartwatch, voice-controlled ovens and refrigerators, a voice-enabled factory-floor assistant, and smart healthcare devices used at home. They illustrate possible application areas, not tested product implementations or a claim that every PSoC Edge variant can support every example unchanged.
Which software tools are used to deploy a voice model?
Infineon describes two related but distinct parts of the workflow. DEEPCRAFT tools cover model and audio work; ModusToolbox is the device-side programming and integration environment.
- Choose a model-development route. DEEPCRAFT Studio is positioned for starting with data collection, preprocessing, training, and deployment. DEEPCRAFT also includes voice-assistant and audio-enhancement solutions that can be customized for wake words and keyword spotting.
- Convert and optimize a model when starting from an existing one. Infineon says DEEPCRAFT Model Converter can accept a model such as one built with PyTorch, then convert, optimize, and validate it for PSoC Edge.
- Integrate it on the device. Use ModusToolbox for programming and device-side integration. The interview describes ModusToolbox and DEEPCRAFT as separate tool families designed to work together.
- Validate the result on the intended hardware. Test recognition quality, response time, memory use, and power for the selected model and device. The interview explains tool roles but does not publish a reproducible deployment recipe or benchmark for a specific model.
Which PSoC Edge board should you use to prototype?
| Hardware option | What the interview says | Useful distinction |
|---|---|---|
| PSoC Edge evolution kit | Described as a full-featured board exposing the family’s broader interfaces | Consider it when you need to explore a wider set of device interfaces. |
| PSoC Edge E84 AI Kit | Described as a lower-cost kit with sensors, microphones, radar, and display connectivity | Consider it when those integrated components fit the prototype you want to build. |
The interview describes the E84 AI Kit as available when recorded, but does not establish current stock, price, or availability by region. Check current distributor listings before choosing a board. Infineon also says designs can start on an E81- or E82-class device for keyword detection and later move to an E83- or E84-class device for more advanced language processing, with software and hardware compatibility across the family; confirm the requirements of a particular design before relying on that migration path.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →What should engineers verify before choosing PSoC Edge?
The interview is a vendor discussion of platform capabilities, not an independent evaluation. Its architecture description can help narrow the questions to test, but it does not supply independent power figures, latency benchmarks, model-accuracy results, comparative pricing, or a reproducible test setup.
- Always-on power: Measure idle and wake-word power with the same model, microphone, and operating conditions used in competing designs.
- NLP fit: Confirm model size, supported operators, quantization options, memory needs, accuracy, and measured response time for the intended task.
- Compute split: Verify that the selected part supports the desired low-power listening path and the accelerator or DSP capabilities needed for follow-on processing.
- Toolchain workflow: Check model conversion, profiling, debugging, data collection, and deployment for the specific model and board.
- System integration: Match audio interfaces, graphics, sensors, radar, SRAM, and connectivity to the product rather than assuming every family member has the same features.
- Security evidence: Confirm the certification and security features relevant to the finished design using documentation for the exact device and implementation.
- Lifecycle and cost: Check current device and kit pricing, availability, software licensing, and long-term support for the intended market.
How to interpret the secure-enclave claim
In the interview, Infineon says PSoC Edge has a secure-enclave architecture and achieved PSA Level 4 integrated secure-enclave certification. Cruz characterized Level 4 as the highest level achieved by a microcontroller. That is the claim as presented in the interview; the episode does not supply a separate certification document. For a security-sensitive design, verify the certification scope and the exact device’s security documentation rather than treating the interview statement as a complete security assessment.
What the EE Times interview adds to the edge-AI discussion
The practical point is not simply that a microcontroller can run a language model. It is that an endpoint can keep a low-power listening function ready, wake more capable processing only after a trigger, and perform some follow-on language work without relying on an internet connection. Infineon’s Cruz described that direction this way: “We are introducing a new paradigm, a new level of processing where you are being actually able to have a natural language processing without relying on the internet connectivity.”
Whether that approach suits a product depends on model quality, measured energy use, response time, available memory, and the degree of offline functionality required. The interview provides Infineon’s architecture and product positioning; those product-level results still need validation for the specific design.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
