Design And Research On Human-machine Speech Interaction System In Embedded Microprocessor

Speech recognition technology has matured through the development in recent years.In a sense,it can be said as the "foundation stone" for the complete realization of "artificial intelligence".At present,the embedded speech recognition system generates the matching model by training a large number of speech commands based on a microprocessor platform(such as chip).And then matching the input speech with the model to complete speaker-independent speech recognition.Although the current research on speech recognition technology has gradually become practical from the laboratory,it will be affected by many objective factors in practical application,such as background noise,which reduces the recognition accuracy and makes the user experience very unfriendly.This paper takes embedded microprocessor as the core,designs and implements a speech control system based on vending machine.This paper first introduces the process of speech signal preprocessing,and adopts a double threshold decision method for endpoint detection.Then the wavelet + MFCC feature extraction algorithm is proposed,which replaces Fourier transform with wavelet transform,and makes up for the disadvantage that the window size of Fourier transform cannot be transformed with frequency.Then,DTW algorithm can effectively deal with the time-varying characteristics of speech.Combining the DTW algorithm with the polynomial kernel function to obtain the PDTW-SVM algorithm.It not only has the ability of local interpolation and global generalization,but also solves the problem of inconsistency between the time-varying characteristics of speech and the equal length vector required by the traditional SVM kernel function.Finally,the PDTW-SVM algorithm is transplanted to the LD3320 platform,and an echo noise reduction circuit based on CX20921 is added according to the actual application scenario to improve the system's noise resistance and stability.In this paper,the speech recognition system is simulated and tested.The experimental results show that the simulated recognition rate of the system can reach 96.49%,and the measured recognition rate can reach 92.17%.It has high practical value.