Deployment of Quantized Deep Noise Suppression on Real-Time Edge Platforms

Key Speaker

Abstract

Many audio devices benefit considerably from noise suppression algorithms, including hearing aids, AirPods, and headsets. In recent years, with the advent of artificial intelligence in several applications, neural network-based denoising algorithms have also become essential in the audio signal processing field. However, many of the previously mentioned devices leverage embedded systems, i.e., architectures with limited computational resources. Therefore, the challenge addressed in this paper is the efficient deployment of neural networks on edge devices while meeting tight real-time constraints. In this paper, we extend our previous work on integrating NeuralCasting with Patmos, providing a WCET analysis and comparing the performance with state-of-the-art AI frameworks (TensorFlow Lite and ONNX Runtime) for deployment on embedded systems and different hardware platforms (RaspberryPi).

Authors

1. Alessandro Cerioli alceri@dtu.dk (Technical University of Denmark)
2. Tórur Biskopstø Strøm tbst@dtu.dk (Technical University of Denmark)
3. Clément Laroche claroche@gn.com (GN Audio)
4. Tobias Piechowiak topiechowiak@gn.com (GN Audio)
5. Luca Pezzarossa lpez@dtu.dk (Technical University of Denmark)
6. Martin Schoeberl masca@dtu.dk (Technical University of Denmark)

 Overview  Program