Voice activity detectors sense if someone is talking and are an integral part of many audio processing applications – from speech recognition in smart assistants to VoIP communication. They differentiate between speech and non-speech audio, in order to trigger the application’s function. For battery-powered devices in particular, the amount of power that VAD requires can significantly impact the size of the device.

“Lowering the power requirements for voice detection will be a key breakthrough that transforms and accelerates the voice technology market,” said Jez Stark, Head of Semiconductor Services. “The growing adoption of voice-based personal assistants such as Alexa™ and Google Home™ are driving the requirement for always-on functionality. At the same time, technology companies are continuing to push boundaries, exploring ways to put voice interfaces into ever smaller devices, such as smart hearables and hearing aids. These dual movements are creating demand for ultra-efficient and ultra-low-power VAD.”

Ecoutez requires just 11µW in 0.2mm2 on a 40nm process. By designing a flexible IP block specifically for VAD applications, Cambridge Consultants is enabling a low power VAD that retains high performance and flexibility. Using the company’s patented Sapphyre™ design flow for custom cores and accelerators, Ecoutez supports low power speech detection while maintaining a flexible architecture. This allows enhancements to the algorithms it runs without requiring a new spin of silicon and without compromising performance or efficiency. The design can also be extended to support keyword detection algorithms.

Sapphyre™ technology applies software principles to the design of custom silicon, enabling fast prototyping, optimization, regression testing and efficient debugging, all of which are essential for minimizing development costs and improving performance.

“While a typical design cycle for a silicon IP accelerator of this complexity would be around 12 months, our Sapphyre™ approach allowed for Ecoutez to be completed in less than 12 weeks,” said Robert Tan, Associate Director. “With this capability, we see opportunities for even further reductions in power consumption, including for a whole variety of customized IP blocks not limited to audio processing.”

編集者へ