BPNN Speaker Model Based on Speaker Characteristic Pattern
Fang Shao · Chinese Journal of Computers · 2002
This paper presents a BPNN speaker model based on speaker characteristic pattern. As a preliminary processing for the BPNN, we use VQ technique to create a common speaker characteristic pattern for all speakers. Using this characteristic pattern, each speaker's sample vectors are mapped to respective membership function vectors, which are served as input parameters for the BPNN. The membership function reflects the average distance distribution between the sample vectors and the characteristic pattern vectors, and bears stronger model discrimination information. By means of mapping, the sample vectors transforms into a one dimension membership function vector, whose length is the size of code books of the characteristic pattern. So it provides a method to normalize the duration for different speaker's utterance and to extract further speaker feature for better recognition performance at the same time.Several experiments are performed under different situations, such as different content of voice samples and different number of speakers, Compared with the traditional methods base on voice incitation algorithm. The results show that the BPNN speaker model based on speaker characteristic pattern bears stronger classification ability and improves the system's recognition performance and practicality accordingly.