End-to-end Keyword Spotting using Neural Architecture Search and Quantization

TL;DR

This paper presents a method combining neural architecture search and quantization to develop efficient end-to-end keyword spotting models that perform well on limited-resource devices, demonstrated on the Google speech commands dataset.

Contribution

It introduces a differentiable NAS approach for CNN architecture optimization on raw audio, coupled with quantization techniques to reduce model size and computational complexity.

Findings

Achieved 95.55% accuracy with 75.7k parameters and 13.6M operations.

Quantization reduced bit-width to approximately 3 bits per activation and weight with minimal accuracy loss.

Compared end-to-end CNN models favorably against traditional MFCC-based systems.

Abstract

This paper introduces neural architecture search (NAS) for the automatic discovery of end-to-end keyword spotting (KWS) models in limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) operating on raw audio waveforms. After a suitable KWS model is found with NAS, we conduct quantization of weights and activations to reduce the memory footprint. We conduct extensive experiments on the Google speech commands dataset. In particular, we compare our end-to-end approach to mel-frequency cepstral coefficient (MFCC) based systems. For quantization, we compare fixed bit-width quantization and trained bit-width quantization. Using NAS only, we were able to obtain a highly efficient model with an accuracy of 95.55% using 75.7k parameters and 13.6M operations. Using trained bit-width quantization, the same model achieves a test accuracy of 93.76% while using on average only 2.91 bits per activation and 2.51 bits per weight.