Sound event detection for low power embedded systems

Abstract: Supervised sound event detection software implementations currently achieve high performance. This allows the development of real-world applications, especially for the growing up domain of the Internet of Objects (IoT). However, current performance is achieved at the cost of hard computational complexity and memory needs, which are inconsistent with low-power hardware limitations. This presentation addresses these limitations in two contexts.

The first part will present the work carried out during my PhD. We consider a very low-power sensor node, inconsistent with neural network implementation. Therefore, only the feature extraction is crafted on the node, and the resulting data are transmitted to a cloud for follow-up processing. In this context, I attempt to understand the effect of both quantization and transmission of features on sound event detection performance.

The second part will address the problem of neural network implementation on embedded systems, that my post-doctorate deals with. A quick state-of-the-art about hardware and software solutions for low complexity implementation will be introduced, as well as discussions about the carbon footprint of current classification algorithms.