Speaker adaptation for Slovak statistical parametric speech synthesis based on hidden Markov models

In this paper, the first experiments with the speaker adaptation for Slovak statistical parametric speech synthesis based on the hidden Markov models are described. The implementation of two new Slovak voices based on the model adaptation technique consisted of small speech corpora design, which were used for the adaptation procedure and the processing of Slovak automatic speech recognition corpora, which were used for the average voice training is described. In the performed listening subjective tests, male and female speech generated from the regularly trained and adapted models was compared with the help of the opinion score and the semantically unpredictable sentences tests. The obtained results showed a high potential of this technique in case of Slovak language, because we manage to get a new intelligible and relatively natural sounding artificial speech with the small amount of data.

I. Introduction

Today, the speech synthesis systems are represented by computer systems which can convert input text into output audio file – speech. The main task of these systems is primarily making life easier for example to people with physical disabilities, such as blind people, or to totally ordinary people and facilitating their day to day operations. The use of speech synthesis in practice is often faced with the reluctance of users to communicate with device, whereas synthetic speech acts artificially and creates a barrier in communication [1]. That is why the research in the field of speech synthesis is focused on the development of new advanced methods and their improvement in order to make final speech output from these systems as similar to the human interpretation as possible.