Speaker Identification Using Discriminative Adapted HCRF models

• Main Authors: Wei-Chen Liu, 劉維宸
• Other Authors: 洪維廷
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/93719537822517049731

碩士 === 元智大學 === 通訊工程學系 === 99 === This thesis presents the techniques of speaker identification using discriminative adapted hidden conditional random field (HCRF) models. A HCRF-based framework is adopted for the universal background model (UBM) and speaker models. We adapt the UBM to form a specific speaker model using a model transformation-based technology in HCRF-based framework with the speaker’s enrollment speech. A novel training algorithm combining the discriminative training criterion with HCRF for speaker identification is also proposed. This work also adopted discriminative training technique to train GMM/UBM, HMM/UBM, and HCRF/UBM speaker models respectively; and the performances of speaker identification by the three speaker models with different amounts of training speech for testing speech were investigated. The experimental results indicate that the HCRF/UBM model consistently achieved the lowest error rate among the three models regardless of the length of the test and training speech and presence of noise.