Low-memory voice keyword detection method and system, medium, equipment and terminal

Abstract

The invention belongs to the technical field of voice signal processing, and discloses a low-memory voice keyword detection method and system, a medium, equipment and a terminal. The method comprises the steps: performing pre-processing, time-frequency domain feature MFCC extraction, attention and time sequence convolution neural network model training on a voice signal; and performing dimension reduction on parameters of a full connection layer in a TACRNN model through the SVD technology, performing low-order quantization on dimension reduction parameters, and reducing the storage amount of model parameters needing to be stored. According to the method, SVD does not need to be carried out on original model parameters to compress the model parameters, the relevance between the model parameters is fully utilized, and the memory requirement for parameter storage is greatly reduced. The method is different from traditional methods that model parameters are represented by double-precision floating points; and on the basis of SVD dimension reduction on the model parameters, the compressed parameters are expressed by low bits, so that the memory requirement on the model parameters is reduced. The deployment capability of lightweight equipment for a voice detection algorithm is further enhanced and adapted.

Description

technical field [0001] The invention belongs to the technical field of voice signal processing, and in particular relates to a low-memory voice keyword detection method, system, medium, equipment and terminal. Background technique [0002] At present: With the development of artificial intelligence, especially the breakthrough in the application of deep learning, the architecture scheme based on deep learning has become the mainstream method of keyword detection. The keyword detection algorithm based on deep learning is applied to keyword detection in the way of deep learning model for the first time, which greatly improves the accuracy of keyword detection, but due to the large number of parameters in the model, it consumes a lot of memory and increases storage costs. ; Based on the deep learning algorithm, the parameters of the model are too large, and the extracted features are redundant, which leads to a large amount of calculation and storage consumption of the model, r...

AI Technical Summary

Problems solved by technology

[0004] (1) The existing technology based on the deep learning algorithm is mainly because the parameters of the model are too large, and the extracted features are redundant, which leads to a large amount of calculation and storage consumption of the model, resulting in an increase in cost, and is not suitable for the application requirements of lack of memory.

[0005] (2) The existing technology is based on the variational feature compression model, which uses Bayesian theory to use the conditional probability P(feature|code) of the code layer and the feature (feature) to approach the conditional probability P( feature|Z), so as to use the code as a new feature, but this method will lose a lot of information about the original feature

Since the data of each cluster has a certain divergence, the relevance of the neural network weight matrix to be compressed is reduced, resulting in unsatisfactory compression performance

[0007] The difficulty of solving the above problems and defects is: the use of deep learning for speech keyword detection has achieved good performance, but it requires large-scale data storage and calculation, making it almost impossible to directly apply it to portable applications

Images