Research Of Automatic Language Identification Of Telephone Channels

Automatic language identification technology is based on phonetics, linguistics, digital signal processing and pattern recognition. It uses computer to analyze and process a piece of speech signal in order to judge which language it belongs to, It is a key important part of the future speech processing flow consisting of automatic language identification, speech recognition, machine translation and speech synthesization.Through the study of the histories of automatic language identification technology, the paper points out that two important problems of features are seeking for the new efficient feature and the balance between high efficiency and the performing cost. In this paper, we make a helpful exploring research by adding the inter-frame dynamic information into the time-frequency principal component (TFPC) feature and introducing the Cohen bilinear time-frequency distribution (TFD) technology into the feature extracting processing.In classical features, most of the parameters are stationary in addition to dynamic differential parameters. But inter-frame dynamic information contains a great amount of language identification information. Through the principal component analysis of the extending feature vectors in time-domain, the TFPC features with inter-frame dynamic information are obtained. The experimental results show that the TFPC is superior to original 12-D LPCC with 10.87%, and the language identification rate of the TFPC feature is 80.18%.Taking the advantages of Cohen bilinear TFD in the nonstationary signal processing, Cohen bilinear TFD are used in this paper to solve the fundamental problems of short-term analysis technique. The extraction of pith frequency is extraction, the improved feature of MFCC, TFMFCC, is obtained. Also a new class-dependent kernel TFD feature is proposed only using Cohen bilinear TFD. The experimental results show that TFMFCC is better than original MFCC with 3.13%, and the language identification rate of the class dependent kernel TFD feature is 54.95%.From the results, we can know TFPCA is good at integrating inter-frame dynamic information into features and the Cohen bilinear time-frequency distribution has the potential to improve the efficiency of the features with the advantages of high time-frequency resolution and being free of short-term stationary assumations.