Study On Speaker Localization And Recognition For An Intra-vehicular Assistant Robot

The intra-vehicular assistant robot is an intelligent vehicle that operates inside the space station to assist astronauts in performing in-orbit missions and carrying out scientific experiments.This paper mainly focuses on the speaker localization and recognition of the assistant robot in the cabin.The delay estimation method based on the fine interpolation of correlation peak is used to realize the speaker localization of the assistant robot,and the deep neural network is used to realize the speaker recognition of the assistant robot.The main work of the dissertation is as follows:The special application background of the space station acoustic environment is relatively stable and the operating range of the assistant robot is limited,so a positioning method based on spatial six-element microphone array is proposed.Considering accuracy requirement of the time delay estimation and computing ability of the platform,an improved time delay estimation algorithm based on the fine interpolation of correlation peak is adopted.The algorithm can break breaking through the limitation of the existing time domain sampling rate,and can effectively weaken the fence effect caused by the FFT transform,thereby improving the correlation function resolution and improving the delay estimation accuracy.Taking into account the computing ability of the robot,the endpoint detection method based on short-term cepstrum distance is used to screen out the effective speech segment and reduce the computational length of the correlation calculation,thus effectively improving the computational efficiency.The environmental noise can cause great interference to the speech quality,which affects the accurate calculation of the delay estimation value.Therefore,the spectral subtraction and quadratic correlation function are added to the algorithm to improve the delay estimation algorithm.The intra-vehicular robot adopts a spherical structure design,so the spatial six-element microphone array system can be used for sound source localization.Compared with the planar array,the design can obtain pitch/ azimuth angle of the sound source(speaker)more accurately,and reduce the influence of delay estimation error on positioning accuracy.Aiming at the problem that the accuracy of the traditional speaker recognition phrase is not high,this paper proposes a speaker recognition method based on the MFCC parameter of speech signal,which adopts the deep neural network as the back-end classifier.The deep neural network has strong multi-layer nonlinear modeling ability,which can deeply excavate the feature parameter information to classify the original speech features.A six-person phrase sound database was established.Through the speaker recognition test,the recognition accuracy rate reached over 95%,which verified the feasibility of the design method.The ground test system of the intra-vehicular assistant robot is established.Experimental results show that the speaker localization method based on FICP algorithm can orientate the sound source target efficiently,and the accuracy of the improved method is better than that of the method based on generalized cross-correlation.The comparative analysis verifies the effectiveness of speaker localization method of the assistant robot.Thus,the method can guarantee the localization needs of the assistant robot.