Evaluation kit offers ML for always-listening edge devices
Article By : Susan Nordyk
Aspinity’s Voice-First Evaluation Kit demonstrates analog machine learning for always-listening edge devices with extended battery life.
Aspinity’s Voice-First Evaluation Kit (EVK2) demonstrates analog machine learning for always-listening edge devices with extended battery life. It features the company’s Reconfigurable Analog Modular Processor (RAMP) chip, which uses near-zero power to analyze raw, unstructured analog microphone data at the start of the signal chain to determine if voice is present prior to triggering the wake word engine.
Since up to 90% of the sound data captured within a 24-hour period is not voice, the RAMP chip’s analyze-first approach minimizes the power-on time of the ADC and wake word engine to increase battery life by as much as tenfold. The device also continuously collects and compresses the 500 ms of sound prior to the wake word (preroll) into approximately 2 kB of memory. Preroll is required by most wake word engines in order to accurately determine that a command has been spoken.
The Voice-First EVK2 provides audio test files for fast startup and a live audio testing option that employs a MEMS microphone from Infineon for design testing. STMicroelectronics’ Nucleo development board outfitted with an STM32H743ZI microcontroller allows testing of analog voice activity detection with or without preroll collection and delivery to a third-party wake word engine. The onboard RAMP chip comes preloaded with voice-activity detection and preroll collection, compression, and reconstruction algorithms.
Aspinity has developed an end-to-end, ultra-low power voice wake-up system using the latest generation of the RAMP chip together with popular MCUs and wake word engines. An application brief is free for downloading on the EVK2 product page.
To learn more about RAMP with VAD and preroll, you can watch this video.
Find more datasheets on products like this one at Datasheets.com, searchable by category, part #, description, manufacturer, and more.
This article was originally published on EDN.
Related articles:
- Audio front-end and wake-word detection: Key considerations in reference designs for Alexa
- Simplifying the deployment of voice detection algorithms for a cloud service
- Voice Recognition to Drive More ‘Conversational’ Platforms
- Multi-microphone reference designs overcome interfering ambient noise
- Alexa, can you hear me now? Low power voice interface technology evolves
- Connectivity remains central to mainstreaming AI, machine learning workloads
