Speech Synthesis System Evaluation

U ovoj disertaciji predstavljena su tri algoritma sinteze, konkatenacija difona, TD-PSOLA i Merlin sustav koji se temelji na neuronskim mrežama. Algoritmi su postepeno složeniji i kvalitetniji jedan od drugog. Napravljena je usporedba u načinu tvorbe zvuka i predstavljene su prednosti i razlike među njima. Kao osnovni cilj ovog rada, napravljena je web stranica za evaluaciju sustava. Korisniku se omogućilo pretvaranje proizvoljnog teksta u govor, preslušavanje i preuzimanje istog. Korisnik sintezu može ocijeniti u kategorijama kvalitete, razumljivosti, prirodnosti i nepravilnosti izgovora. Korisnik također ima uvid u sve do sada sintetizirane rečenice na web platformi, uz njihove ocjene i postotak biranja pojedinog algoritma. Također, korisnik ima mogućnost ocjenjivanja algoritama na globalnoj bazi za koje je upravitelj web stranice pripremio par snimki za pojedini algoritam.his disertation presents three algorithms for speech synthesis, diphone concatenation, TDPSOLA, and a neural network-based Merlin system. Algorithms are gradually more complex and of better quality than each other. A comparison is made in the way sound is formed and the advantages and differences between them are presented. As the main goal of this paper, a website for speech quality evaluation was created. The user was allowed to convert arbitrary text into speech, listen to it and download it. The user can evaluate the synthesis in the categories of quality, intelligibility, naturalness and irregularity of pronunciation. The user also has an insight into all the sentences synthesized so far on the web platform, along with their ratings and the percentage of selecting a particular algorithm. Also, the user has the ability to evaluate algorithms on a global basis for which the website manager has prepared a couple of snapshots for each algorithm