Abstract:Keyword spotting (KWS) is one of automatic speech recognition (ASR) research fields. KWS algorithms based on neural network are remarkably better than others. It is extremely suitable for KWS algorithms to deploy on portable hardware devices to achieve excellent performance, but the number of weights is too large to apply due to limited hardware storage resources. Quantization is a natural approach to solve this puzzle. We propose a quantization formula to deploy KWS algorithms based on neural network on hardware devices effectively. In this paper, we consider symmetry of weights about the origin, and reorganize the order of 1-bit quantization of residual error and alternating quantization of minimization, proposing our translational bit-by-bit multi-bit quantization. Furthermore, we implement our method to convolutional recurrent neural network (CRNN) for KWS problem, analyzing results on two aspects, root mean square error (RMSE) and accuracy respectively. We can obtain smaller RMSE between quantized weights and full precision weights, the cost of it is the increment of computational complexity, but we can execute quantization in every mini-batch to constrain training time within a reasonable range. Our formula can get an acceptable accuracy by 3-bit quantization, with ~10.65× memory saving. And we can obtain no loss accuracy comparing to full precision by 4-bit quantization, sometimes even better, with ~7.99× memory saving. Our quantization method can achieve better performance and save memory resources, which are able to be used to deploy KWS algorithms based on neural network on hardware devices effectively.

Convolutional Recurrent Neural Networks for Small-Footprint Keyword Spotting

Low-Latency Convolutional Recurrent Neural Network for Keyword Spotting

Small-Footprint Keyword Spotting with Multi-Scale Temporal Convolution

A Separable Temporal Convolution Neural Network with Attention for Small-Footprint Keyword Spotting

Small-footprint Keyword Spotting with Graph Convolutional Network

Focal Loss And Double-Edge-Triggered Detector For Robust Small-Footprint Keyword Spotting

Hello Edge: Keyword Spotting on Microcontrollers

A Depthwise Separable Convolution Neural Network for Small-footprint Keyword Spotting Using Approximate MAC Unit and Streaming Convolution Reuse

Frequency & Channel Attention Network for Small Footprint Noisy Spoken Keyword Spotting

Depthwise Separable Convolutional ResNet with Squeeze-and-Excitation Blocks for Small-footprint Keyword Spotting

Low-complex and Highly-performed Binary Residual Neural Network for Small-footprint Keyword Spotting

Hierarchical Neural Network Architecture In Keyword Spotting

A Monaural Speech Enhancement Method for Robust Small-Footprint Keyword Spotting.

Predicting detection filters for small footprint open-vocabulary keyword spotting

A 3.8-Μw 10-Keyword Noise-Robust Keyword Spotting Processor Using Symmetric Compressed Ternary-Weight Neural Networks

Keyword Spotting for Hearing Assistive Devices Robust to External Speakers

Translational Bit-by-Bit Multi-bit Quantization for CRNN on Keyword Spotting.

Effective Combination of DenseNet and BiLSTM for Keyword Spotting.

Neural Networks for Keyword Spotting on IoT Devices

Separable Temporal Convolution plus Temporally Pooled Attention for Lightweight High-performance Keyword Spotting

Text-Dependent Speech Enhancement for Small-Footprint Robust Keyword Detection