A Novel Method for Rapid Speaker Adaptation Based on Support Speaker Weighting
Tie Cai, Jie Zhu · 2006
In this paper, we propose a novel model-based speaker adaptation method called support speaker weighting (SSW), which performs the adaptation scheme of model combination based on the selected speakers. These speakers, who are acoustically close to the test speaker, are selected from reference speakers using support vector machines (SVM). Compared with GMM/HMM based speaker selection methods, the proposed method can quickly obtain a more optimal speaker subset because the selection is dynamically determined according to the distribution of reference speakers around the test. Experimental results for a large-vocabulary task given in this paper show that this method is both cheaper in terms of memory and more effective than reference speaker weighting (RSW) for tiny amounts of adaptation data. Relative error rate reduction of 4.1% is achieved when only one adaptation sentence is available.