Fraunhofer EMFT in SOIL
As leader of work package 3 in the SOIL project, Fraunhofer EMFT coordinates the development of critical FDSOI-based IP blocks, which serve as building blocks for later system demonstrators.
Within this work package, Fraunhofer EMFT also develops an ultra-low-power Keyword Spotting (KWS) analog front-end implemented in GlobalFoundries 22FDX technology. The concept, however, is technology-agnostic and can be transferred to any CMOS-platform where energy-efficient analog processing is required.
Why ultra-low-power KWS?
Many modern devices require always-on acoustic awareness, for example:
- Voice-activated user interfaces
- Smart home assistants and IoT devices
- Battery-powered sensors in wearables or environmental monitoring
- Devices that must remain alert without draining the battery
Traditional digital KWS systems rely heavily on ADCs and DSP blocks, which consume significant energy even when no keyword is spoken. In contrast, a neuromorphic analog front-end boasts with features which make true always-on capability feasible:
- Processing sound in an event-driven mannter
- Extracting only the essential acoustic features
- Significantly reduced data rate and computation
- Multi-year battery operation or energy harvesting
These capabilities are key for ultra-low-power edge devices, which must listen continuously without compromising battery life.
Technological highlights
- Analog feature extraction inspired by cochlear mechanisms
- Conversion to spike-based representation for Spiking Neural Networks
- Fully asynchronous operation for high temporal accuracy
- Targeted power consumption in sub-µW for always-on applications
- Designed in 22FDX with architecture being transferable to other technology nodes
Dissemination and early results
Presentation of SOIL project WP3 activities on IP-SoC 2024 Conference (Fraunhofer EMFT / Damian Panter):