Design And Implementation Of Low Delay Terminal Command Word Recognition System Based On Deep Learning

Speech recognition system as an important part of intelligent human-computer interaction technology has a wide range of applications in various fields.Continuous speech recognition technology generally has complex model,large demand for calculation,and difficulty in terminal recognition.At present,it mainly adopts online network recognition,which has high cost and large delay.For some control terminal devices,a dozen to a few hundred key command words can cover all functions,in which case command word recognition is more appropriate than continuous speech recognition.But the current command word recognition system exists to support the command word number is less,the recognition rate is not high,the system complex,high identification delay,power consumption big problems,in order to solve these problems,the paper designs a new type of low delay deep learning can be applied to terminal command word recognition system,and completed the realization of the terminal hardware platform.Aiming at the problem that the command word recognition feature needs to simplify and accurately express the speech information,the mel filter bank output feature is adopted as the input feature of the recognition system.In this paper,the process of speech signal processing and the requirements of deep learning network input characteristics are analyzed in detail.The speech power spectrum is filtered by nonlinear frequency mapping to the filter bank constructed in the mel domain,so as to retain the useful information of low frequency and reduce the redundant information of high frequency.Experiments show that this feature can express the content information of the command word well and reduce the computation of the recognition network.To solve the problem of matching the relationship between audio frames and pinyin labels for command word recognition,long short-term memory recurrent neural network was used to model the problem.This paper detailed analyzed the deep neural network structure,convolution neural network structure and recurrent neural network,by using the LSTM recurrent neural network nonlinear activation function,the structure of the gating time dimension parameters of historical status,and network sharing mechanism,improve the model for complex speech distribution fitting ability,meet the processing requirements of variable length sequences,can make the model with the speech of the input frame by frame processing in order to reduce delay.The experimental simulation results show that the model has a high accuracy rate for the recognition of the test set,and the test on the hardware platform shows that the recognition delay of the model is very low.A maximum probability matching confidence algorithm is proposed to solve the problem of probability graph transformation of deep learning recognition model.The algorithm uses the information in the probability graph to calculate the maximum probability of the possible occurrence of each word in this speech.After comparison,the command word corresponding to the maximum value is selected as the result.Besides,the accuracy and error rate can be balanced by setting the threshold value and the maximum value comparison.This algorithm not only accurately identifies the target command word,but also reduces the probability that other speech input is misidentified as the target command word.In order to verify the actual application performance of command word speech recognition system,this paper combined with STM32F407 terminal hardware platform to design the whole block diagram of the command word recognition system,in the light of the features of the terminal hardware platform to implement and accelerate the algorithm,and the deployment to the terminal for each algorithm module of the hardware platform for validation.The verification results show that the system can run completely on the terminal hardware platform with low computing resources.In addition,the data recorded by the terminal hardware platform is tested,which proves that the command word recognition system based on deep learning has high recognition rate and low delay in practical application.