Abstract
A text-dependent speaker recognition method using trapezoid fuzzy similarity function to decide how similarity between voice features of a test user with those of the registered speaker with nearest distance is proposed. The trapezoid fuzzy similarity function is constructed based on three-time data recorded during enrolment to present personal identification voice (PIV) and a statistical data of a person recorded many times in a long period to absorb the intra-variation. A set of acoustic voice features is also introduced to present some general speaker and text dependent characteristics that are effective for modeling PIV, thus can capture the inter- variation from one speaker to another. The experimental results on 24 speakers recorded in 4 different sessions show that, without false accepting, the proposed system can decrease 23.15% of false rejecting cases, compared to traditional nearest neighbor approach. The focus of this work is on applications which require fast processing and few burdens for users.
