Neural Architecture Search for Efficient Mixed-Precision Quantization in Audio Denoising

This work presents a neural architecture search approach based on evolutionary algorithms to tackle the problem of mixed-precision quantization considering 8, 16, and 32-bit depths for NSNet2, a neural network used for audio denoising tasks. The goal is to preserve high network performance while reducing the model size, making it suitable for deployment on embedded systems such as headsets. The neural architecture search (NAS) algorithm considers metrics such as (a) Perceptual Evaluation of Speech Quality (PESQ) to assess the audio quality after de-noising, (b) the inference time of the neural network, and (c) its memory footprint. Our results showcase remarkable improvements, with a 29.93% reduction in inference time and a 57.88% reduction in memory footprint compared to the full-precision float32 baseline, while preserving high performance in audio de-noising with a negligible drop of PESQ metric. This work provides a framework for estimating the optimal mixed-precision quantization configuration, which is beneficial for deploying neural networks on resource-constrained devices.