Speaker Recognition via Statistics of Acoustic Feature Distribution

In recent years, with the widely application of speaker recognition, besides recognition precision, people pay more attention to processing speed. The traditional recognition frame which directly uses acoustic feature sequence to match the model of target speaker doesn´t work well in real time environment. So new recognition frame based on the statistic of acoustic feature sequence has arisen. By using acoustic feature distribution around common codebook to model speaker´s charactistic, a new speaker recognition algorithm is proposed in this paper. The common codebook is generated via the training data from all reference speakers, which is used to classify speech feature space, and the model of each reference speaker is described by the statistics of speaker´s acoustic feature distribution around the common codebook. In the recognition, pairwise sequence alignment is adopted to measure the distortion between the acoustic feature distribution of test speech and each reference speaker model, and recognition result is derived by distortion comparison. Experimental results showed that the method proposed in this paper can save calculation and space resource while having better performance over current algorithm which also is based on the statistic of speaker´s acoustic feature distribution.