Edge computing-oriented lightweight voice keyword recognition method

Abstract

The invention discloses an edge computing-oriented lightweight voice keyword recognition method, which comprises the following steps of: preprocessing a signal, and removing a noise signal; performingacoustic feature extraction; constructing a lightweight speech keyword recognition model EdgeCRNN adopting a first-layer feature enhancement method and a lightweight component, wherein the lightweight component comprises a depth separable convolution and residual structure; constructing a basic module Base-Block and a down-sampling module CRNN-Block suitable for a voice keyword recognition task,and constructing an EdgeCRNN based on the basic module and the down-sampling module; and inputting the features into an EdgeCRNN model to carry out speech recognition. According to the method, a lightweight voice keyword recognition model is designed by adopting a feature enhancement method, depth separable convolution and a depth residual structure, so that the hardware resource consumption is greatly reduced, the model can stably and smoothly run on resource-limited equipment, and the leakage of user privacy is avoided.

Description

technical field [0001] The invention belongs to the technical field of speech recognition, and in particular relates to a lightweight speech keyword recognition method oriented to edge computing. Background technique [0002] Speech keyword spotting (Keyword Spotting, KWS) applications usually use the mode of terminal data collection and cloud server recognition. Although the cloud server has sufficient storage space and powerful computing power, it can store and process a large amount of data, but this mode has potential delays. Moreover, with the rapid growth of data, the pressure on the server to process data and the network bandwidth consumed by transmitting data will increase exponentially, which puts forward higher requirements for server computing power and network bandwidth, and the delay will also become more severe. Big, this will be a very bad user experience for applications based on the KWS model. In addition, when user data is uploaded to the cloud server, the...

AI Technical Summary

Problems solved by technology

This model not only has high production costs, but also has the problem of user privacy leakage

[0006] In addition, the existing technology has serious limitations in the application of edge devices. This is because the performance of hardware such as the CPU and memory of the edge devices is relatively poor. For example, the CPU of the Raspberry Pi 3B+ edge device is a single-core ARMs7I chip. , the main frequency is only 1.2GHz

However, the inference process of the deep learning model is particularly demanding on computing power. When running the deep learning model on the edge computing device, there are often situations such as freezes, full CPU usage and shutdown, etc., which limits its application on the edge computing device. practicality

Images