Objective Evaluation Of Speech Quality Based On Stacked Auto-Encoder

In the process of development and training of communication countermeasure equipment,it is important to evaluate the interference effect on the speech communication.This paper studies the objective evaluation of speech quality because the subjective evaluation is time-consuming and arduous.The main work and innovation points are as follow.(1)Building a speech databaseThis dissertation constructs the original speech database and designs the distortion system.The distorted speech set is obtained and evaluated subjectively.Then the distorted speech database with DMOS label is formed.(2)Speech segment extraction based on information energy ratioThe speech signal is pre-weighted,framed and windowed,and the speech activity detection is based on information energy ratio then the speech segment is extracted.The experimental results show that the algorithm can reduce the number of frames of speech endpoint detection.(3)Feature parameter extraction based on Rasta filterThe Rasta filtering is conducted based on the work of Mel filter,and then the RMFCC is extracted.The Rasta filtering is conducted based on the work of Gammatone filter,and then the RGFCC is extracted.(4)Quality judgement based on Stacked Auto-EncodingBased on Stacked Auto-Encoder,the internal structure of speech characteristic parameters is researched by iterative of pre-training to reduce the reconstruction error,and the softmax classifier is combined with SAE to produce quality discriminator to judge the quality of speech.Three tests are set to confirm the effectiveness of the above method.(5)Performance testBased on the correlation coefficient between subjective and objective DMOS,two test of evaluation performance are conducted respectively to confirm the improvement on speech quality using the speech segment extraction based on information energy ratio method and feature parameter extraction method based on Rasta filter...