Within ANDANTE, SynSense develops real-time audio processing and monitoring use-cases for smart-home and battery-powered environments. SynSense deployed acoustic environment and audio event monitoring applications on their audio inference chip Xylo™, running continuously at < 500μW dynamic power. This will enable the deployment of ultra-low power audio processing applications on consumer devices.
The platform developed by SynSense (AFE + Xylo SNN) is used in ANDANTE project in two use cases:
- Use case 3.2 Underwater acoustic signal classification described in this newsletter in the previous content #2.
- Use case 5.1 Consumer auditory processing, which includes the following three subcases:
- Audio event detection for home security: This subcase, used in smart home speakers, is a continuous audio monitoring use case, where distinct classes of audio events such as glass break (for security purposes), distress call (for health monitoring purposes), baby cry, dog bark, etc., are recognized.
- Audio scene classification for consumer headsets: This subcase classifies the noise environment surrounding the user automatically and continuously.
- Microphone array processing for source localization: This subcase, used in smart home speakers, allows the localization of spoken audio with high accuracy. It is a common approach for assisting in noise reduction. This process is often called “Reverse Beamforming”.
As part of the ANDANTE project, SynSense has developed an audio front-end IP complementary to its Spiking Neural Network (SNN) inference core, combined in an SOC “Xylo™ Audio”. Together this provides a complete solution for advanced audio inference, as shown in Fig. 1.
The audio front-end (AFE) is a low power analog design (< 100 μW) tailored for human speech applications supporting an input bandwidth of 100 to 20 kHz, and performs analog filtering through a set of tuneable bandpass filters for further processing by an SNN.
As an active participant in ANDANTE, SynSense recently presented their latest achievements in the Xylo™ family of devices. This includes demonstrating seven classes of ultra-low power and low-latency applications, accompanied by demos highlighting their capabilities.
The Xylo™ family of devices enable real-time always-on temporal signal processing and recognition applications at ultra-low-power (<500μW). Combined with a range of efficient direct sensor interfaces, Xylo™ is ideally suited for intelligent processing of a variety of signals, including MEMS microphones, temperature sensors, pressure sensors, vibration sensors, acceleration sensors, gyroscopes, EMG, ECG, and more.
Xylo™’s SNN core simulates up to 1000 leaky integrate-and-fire (LIF) spiking neurons for configurable and flexible inference applications.
For additional information, contact: dylan.muir@synsense.ai