Eigenspace-based speaker adaptation methods in Persian speech recognition systems

Among speaker adaptation algorithms, eigenvoice (EV) and eigenspace-based MLLR (EMLLR) adaptation approaches have been proposed for rapid adaptation with very limited adaptation data. In these methods, a speaker adapted model is constrained to be a weighted combination of some orthogonal basis vectors. In this manner, both the number of parameters to be estimated from the adaptation data, and the required adaption data dramatically decrease. Although these two algorithms have an acceptable performance for adaptation data in the range of 5 to 10 seconds of speech wave, availability of a large amount of adaption data does not necessarily lead to more efficient models. Experimental results of applying EV and EMLLR adaptation algorithms on FARSDAT database discussed in the paper show that by a limited supervised adaptation data (5-10 seconds), these methods lead to respectively 5.9% and 5.3% improvement in phoneme recognition rate. Furthermore, they yield about 4% improvement in unsupervised adaptation, where the common speaker adaptation methods such as MLLR, cannot work efficiently through a limited supervised or unsupervised adaptation data. In addition, in this paper, the development of EV performance in a large amount of adaptation data is achieved by segmenting the eigenspace based on model characteristics. Keywords-speaker adaptation; principal component analysis; eigenvoice; eigenspace