LSTM Recurrent Neural Network Speech Recognition System Based on Genetic Algorithm Improved HMM
Year: 2025
Type: conference-paper
Abstract: Speech recognition technology has had a profound impact on human life since its inception and is a hot research topic in academia. This article focuses on an LSTM recurrent neural network speech recognition system based on genetic algorithm to improve HMM. The experimental network model adopts a recursive neural network with a layer of 1024 dimensional storage units and gates, combined with stochastic gradient descent method. When collecting, irregular samples affected by accidental factors are removed, and a fixed frequency of 16KHz is used for sampling and quantization to ensure speech quality without occupying memory. In pre training, although RBM pre training is not critical when the amount of data is large, it can optimize the initialization depth, which is beneficial for genetic algorithms. Normalize feature parameters by compressing speech frames to align them with the network vector dimension. In the speech recognition stage, first train the monaural GMM-HMM model and force sequence alignment, then train the Tri1 model for refinement, using linear discriminant analysis for dimensionality reduction and feature-based maximum likelihood linear regression for feature normalization. The experiment showed that the recognition rate was highest at 100%, lowest at 85%, and average at 91.5%, indicating that LSTM significantly improves the recognition accuracy of the system.
Authors: Qingya Hu, Kang Le, Zihao Zhou, Yan Tang
Institutions: Wuhan Donghu University, Central China Normal University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413812470