Computing and Informatics

This paper describes the development of a Slovak text-to-speech system which applies a technique wherein speech is directly synthesized from hidden Markov models. Statistical models for Slovak speech units are trained by using the newly created female and male phonetically balanced speech corpora. In addition, contextual informations about phonemes, syllables, words, phrases, and utterances were determined, as well as questions for decision tree-based context clustering algorithms. In this paper, recent statistical parametric speech synthesis methods including the conventional, STRAIGHT and AHOcoder speech synthesis systems are implemented and evaluated. Objective evaluation methods (mel-cepstral distortion and fundamental frequency comparison) and subjective ones (mean opinion score and semantically unpredictable sentences test) are carried out to compare these systems with each other and evaluation of their overall quality. The result of this work is a set of text to speech systems for Slovak language which are characterized by very good intelligibility and quite good naturalness of utterances at the output of these systems. In the subjective tests of intelligibility the STRAIGHT based female voice and AHOcoder based male voice reached the highest scores.