Neural Network Chips Bring Voice Activation to Battery Powered Devices

The universe of voice-enabled devices could be expanding, thanks to recent advances in ultra-low power microcontrollers that use deep learning to recognize common wake words. The Syntiant NDP100 and NDP101 neural decision processors (NDP) consume orders of magnitude less power when listening and provide 20 time more throughput than typical microcontroller units, according to Syntiant. The result: Battery-powered devices like Bluetooth headphones can be outfitted with always-on voice activation.

The Syntiant NDP100 and NDP101 chips in August passed Amazon's Alexa qualification program, making them eligible for use with the Amazon Alexa Voice Service (AVS). The approval means OEMs can integrate the chips into power-limited portable devices, enabling hand-free interactions with Alexa and taking advantage of new features and Alexa skills.

"Syntiant's neural decision processor technology takes a unique approach to wake word detection on voice-enabled devices," said Sanjay Voleti, senior manager of device enablement with the Alexa Voice Service. "We recently conducted acoustic performance tests of their latest solutions, and were impressed by their ability to support compute-on-edge and wake word detection on low-power devices."

The NDP100 and NDP101 chips consume just 150 microwatts while listening for work words. As Jeremy Holleman, chief technology officer of Syntian described in a blog post, the purpose-built compute engine for neural inference in the NDP series chips provides dramatic energy savings when compared to the stored-program architecture used in traditional microcontroller solutions.

Kurt Busch, chief executive officer of Syntiant, said: "These chips are purpose-built for keyword spotting such as wake words like Alexa, and now our processors can be used for quickly developing voice applications in battery-powered devices."

About the Author

Michael Desmond is an editor and writer for 1105 Media's Enterprise Computing Group.