End-to-end Keyword Spotting using Neural Architecture Search and Quantization
This work addresses the need for compact and accurate keyword spotting systems in resource-constrained devices, representing an incremental improvement through automated optimization and compression techniques.
This paper tackled the problem of designing efficient keyword spotting models for limited-resource environments by using neural architecture search and quantization, achieving 95.55% accuracy with 75.7k parameters and 13.6M operations, and 93.76% accuracy with low-bit quantization.
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.