Title: Speaker Adaptation of Hidden Markov Models Using Maximum Likelihood Linear Regression. Author: Supervisors

Material and results from the current thesis may be used freely provided that the source is stated. Abstract The work presented in this report focuses on an essential problem when doing speaker adaptation; namely how eeectively the speaker speciic information in the adaptation data is used. In the project a system has been implemented for speaker adaptation of hidden Markov models (HMM's) using the Maximum Likelihood Linear Regression (MLLR) method. MLLR is a method that transforms mixture components of HMM's by multiplying the mean vectors with a transformation matrix. It introduces the concept of regression classes as a set of mixture components that are transformed similarly. The adaptation technique is implemented in C. The data used in the tests are taken from the Danish EU-ROM.1 database. All results are averaged over ten speakers. Three issues have been addressed: 1) the eeect of varying the amount of adaptation material, 2) the eeect of using diierent regression class divisions and 3) the importance of the phonetic content in the adaptation material. Tests show that the MLLR technique is very data eeective. Only 3s of speech is needed when a diagonal transformation matrix is used before a positive eeect of the adaptation is seen. When using a full matrix 5s are suucient. It was observed that there is a high dependence between the achieved performance and the regression class division. Transforming each mono-phone individually improves the phoneme cor-rectness from 50.9% for the initial speaker independent models to 58.8% for the adapted models. Based on several approaches it was concluded that there are diierences in the speaker speciic information available from diierent phonemes. Vowels were seen to vary a lot from speaker to speaker and to have relatively much innuence on the eeect of the adaptation. Fricatives on the other hand have very small inter speaker variances. The project addresses the problem of speaker adaptation of hidden Markov models using the Maximum Likelihood Linear Regression technique. Some prior knowledge of the theory of Markov models is assumed. The report is divided into four parts. The Preliminaries part contains the introductory sections including the deenition of the project. In the Theory part theory, techniques and methods are described. The implementation and tests are described in the Implementation, test and conclusion part. The last part of the report is the Appendix where additional information is given. Developed software is included on a oppy disk. References to …