Sonar V2.0: An Open Platform For Speaker Recognition

Nowadays, novel Speaker Recognition (SR) technology and algorithms come forth at all times, which become a great challenge for research and experiment in this research field.This thesis compares and investigates several related existing systems, especially our former Speaker recOgnitioN softwAre platfoRm version 1.0 (Sonar v1), finally the next generation system Sonar version 2.0 (Sonar v2) is implemented.Sonar v2 platform supports multi-level speaker recognition research and development, which can be used for teaching as well as for research. It offers a complete process of speaker recognition, including signal collection, preprocessing, feature extraction, modeling, score optimization and evaluation; what's more, it integrates the most popular and state of arts algorithms in each module. Sonar v2 outweighs Sonar v1 in framework design, usability, extensibility, space & time effectiveness, robustness, documentation, etc. Compared with Alize, which is a famous platform in this field, Sonar v2 supports more feature extraction and modeling algorithm, it shows advantages in system synthesis and accuracy, what's more, the work for teaching platform construction is also a plus for Sonar v2.This thesis mainly focuses on following work:1. Design Sonar v2 framework, implement and validate the system functions: Sonar v2 is provided with clear module partition, specific file definition, transparent underlying file system and scripts based automatic audio corpus evaluation.2. Analyses and improve the space & time effectiveness and robustness of the system: Sonar v2 takes in the advantages and overcomes the disadvantages of existing systems, especially Alize and Sonar v1, which helps it gain better space & time effectiveness and robustness.3. Baseline evaluation: evaluate the performance of implemented algorithms, and make sure the baseline experiments are repeatable.4. Teaching platform construction and system promotion: Sonar v2 comes up with a complete promotion and spread plan, including programming specifications, API documents, client usability survey, etc. Sonar Application and Sonar Practice system are also fulfilled to meet the requirements of teaching. The work forteaching platform construction makes Sonar v2 a practical platform for teaching and fill the blank of the teaching platform in speaker recognition field.The work is supported by the following foundings: NCET-04-0545, NSFC₆0525202/60533040, 863 Program₂006AA01Z136, PCSIRT0652, ZPNSF_Y 106705.