Research And Implementation Of Speech Separation Technology

Speech separation technology,as an basis for speech synthesis and recognition,plays an important role in the field of speech signal processing technology.The traditional single microphone speech separation method based on ideal near-field environment without noise and reverberation can separate the mixed speech signals better,but its separation effect in multiple sound source or large noise environment is not good enough.Based on the microphone array speech separation algorithm by using the beam forming method of target sound source signals to obtain a higher gain and on the non-target direction were strong restrain,to obtain better speech separation performance.But for speech,the obvious bandwidth and instability of the specific causes of the speech separation method in the acquisition of signal relevance difficulty is far difficulty than the traditional antenna array in a stationary narrowband electromagnetic wave signal.Therefore,the paper focuses on how to obtain the correlation of the speech signal,and makes a more detailed analysis and improvement on the existing speech separation method.In this paper,the implementation and improvement based on a dual microphone array speech separation method,and the simulation design and implement based on a spherical tetrahedron microphone array speech method.This paper based on ICA single grain speech separation method,and the PESQ speech quality evaluation method is evaluated by the above two algorithm to separate the speech quality.The results showed that the approved method can achieve better speech separation performance.This paper focuses on the realization and improvement of the speech separation system based on two microphone arrays.In detail,the main work is as follows:First of all,the paper gives a comprehensive introduction to the principle of speech separation based on microphone array.The characteristics of speech and noise and the difficulty of speech signal processing in array are described in detail.Starting from the wave equation in the far-field wideband,signal model is involved in the derivation of microphone array.And the common microphone array topology were introduced,and MVDR beam forming method and FIR filter of the broadband beam forming method based on the principle of a simple description,laid the theoretical foundation for the next step of the system implementation and improvement.Secondly,the speech separation method based on the dual microphone array is realized.The system mainly consists of three parts,namely,the speech activity classification module,the speech separation module and the rear examination module.The speech activity module for speech activity automatic identification of sound source is active,and will send the results to the automatic control assembly speech separation module,so as to control the on-off state of the MVDR beamformer adaptation,so as to obtain the correct correlation of speech signals;speech separation module for mixed speech signals the microphone array received accurate separation,and in order to avoid the signal output of the phase discontinuity,module selection MVDR beamformer and FIR filter combined method to separate the wideband speech signal;post inspection module classification of speech activity on previous results were examined and corrected by the power of the output signal.In order to obtain more accurate results of speech separation.At the same time,the correlation window value and threshold value are theoretically deduced,and the conclusion of SAC module and PC module is proved.Thirdly,this paper designs and realizes a speech separation method based on spherical regular tetrahedral microphone array.The principle of which is similar to the above two microphone array system,the difference lies in the method using harmonic domain applied in the treatment of mixed speech signal with noise and its advantage lies in the harmonic domain of weight vector and cross power spectrum matrix calculation as well as the array manifold matrix form compared to element space is more simple.In the last part,the paper based on ICA single grain speech separation method as the reference,using the PESQ speech quality evaluation method to evaluate the three separation algorithm of speech quality.The results showed that separation method using the design of the speech signal PESQ score mean and variance are high in the former two methods,reflecting design method can achieve better speech separation performance.At the end of the paper,the advantages and disadvantages of the design method are summarized,and the further improvement direction is proposed.