Acoustic-Based Research On Automatic Language Identification

This paper focuses on the research of Automatic Language Identification based on GMMs utilizing several kinds of acoustic features. Main research work includes the following:1. Building up a multiple lingual corpus including speeches of Chinese, English, German, Japanese, French and Spanish. The training set contains speech pieces from 264 speakers with duration from 60s to 300s. Fifty speech utterances with average duration of 4.5 seconds from 30-50 speakers outside the training set are chosen for each language to form the testing set.2. Primary experiment on ALID systems. GMM systems for the task of six-language recognition are established in order to investigate the relations among recognition rate, number of the GMM mixtures and quantity of the training data. The effect of RASTA and CMS on the performance of MFCC is also discussed.3. A new quadratic warping function is proposed for speaker normalization employing pitch mean based frequency warping. Comparative test with traditional linear function and piecewise linear function is performed to verify the validity of the new quadratic warping function.4. Shifted cepstra (SC) is proposed and applied to both RASTA-MFCC and RASTA-PLP in comparative tests with the prevailing SDC features. Effects on performance of different parameter configurations N-p-k for RASTA-PLP-SC are studied and the best performing path is determined through the hill-climbing method. Data fusion of RASTA-PLP-SC with other forms of vector is investigated at both feature and decision level.