Auditory-Acoustic Hybrid LPC Study

Linear prediction analysis plays a central role in speech coding technology, perceptual weighting filter, synthesis filter and logarithmic gain filter all use linear prediction analysis. The main work of this paper is to present a new type of LPC called Auditory-Acoustic- Hybrid -LPC, which combines both acoustic and auditory characteristics, making improvement to the audio quality of synthesized speech and increasing PESQ value. Experimental results show that the H-LPC can improve speech quality of coding algorithm, and it plays a positive role on the encoding algorithm.Firstly, the solution of the H-LPC algorithm is studied. Based on the principle and algorithm analysis of Mel cepstral coefficients (MFCC) and LPC Mel cepstrum coefficients (LPCMCC), this paper integrates MFCC's features with LPCMCC's features and use MFCC to adjust LPCMCC in order to enhance LPCMCC's auditory characteristics which later has feedback to the LPC. The aim of integrated algorithm is to add auditory characteristics to LPC while the acoustic characteristics of LPC keep the maximum extent.In order to detect degree of new features added and loss of acoustic characteristics on H-LPC, experiments has been done using power spectrum density (PSD) estimation testing. The PSD algorithm of original voice uses the direct method. PSD algorithm of LPC chooses Yule-Walker estimation of AR model. The results show that compared with traditional LPC, the PSD of H-LPC reduces, to some extent. However, H-LPC has basically reached the current requirements and the embodiment rate of PSD for the original speech is approximate 95%.Before H-LPC adding into coding algorithm, the first step is to unify various parameters of H-LPC algorithm and synthesis filter algorithm in coding. Priority is to join the hybrid window. Through comparative study of the MFCC and LPCMCC algorithm, it can be found that MFCC can directly use a hybrid window instead of Hamming window while produce no superposition and recursive effects on the process; LPCMCC involved superposition and recursive operations of autocorrelation function, this part of the computing process has been correction. After this part of work has been finished, join H-LPC into 8Kbit / s LD-aCELP coding algorithm.Listen and test with PESQ to synthesized speech obtained by 8Kbit/s LD-aCELP speech coding algorithm which is added H-LPC. Research shows that the audio quality of synthesized speech of it has almost the same level of G.728 coding algorithm, and higher than 8Kbit/s LD-aCELP speech coding algorithm. The PESQ value of LD-aCELP algorithm adding H-LPC is 3.630.