Specific acoustic unit processing in concatenative Romanian speech synthesis used for talking agents

In this paper we present a number of operations involving the manipulation of acoustic units necessary for creating a Romanian speech synthesis database. We propose a recording technique based on the newly introduced concept of the checkup word, we develop an acoustic unit extraction algorithm and we advance a set of identification markers. Also we propose a simple and efficient sound level normalization method based on the checkup word, which proved to have similar results when compared to a sophisticated method based on automatic gain control. Moreover, the architecture of the Romanian speech synthesis module, also detailed in this paper, is integrated with a Romanian-specific 3D visual speech simulation model resulting the very first application of a Romanian 3D talking head. Preliminary listening tests revealed a 4.5 mean opinion score on a five point Likert scale regarding the synchronization performance of the audiovisual stream.