Objektivno vrednotenje sistemov za umetno tvorjenje govora

MULEJ, MARKO (author), Mihelič, France (mentor)

Namen naloge je oceniti možnost uporabe objektivnih evalvacijskih metod za ocenjevanja kakovosti umetno tvorjenega govora. Pomemben korak v razvoju sintetizatorjev umetnega govora je ugotovitev kakovosti. Trenutno so za njeno ocenjevanje najboljše subjektivne metode, vendar so te pri njihovi uporabi časovno in denarno potratne. Predvsem za razvijalce je uporabno, da so na voljo načini, s katerimi se lahko hitro in enostavno preveri kakovost samega stroja in različnih načinov kalibracije avtomata. V okviru te naloge sem uporabil dve računski metodi evalviranja umetnega govora, Mel-kepstrovo razdaljo in DTW (metoda dinamičnega ukrivljanja časa), na štirih različnih sistemih za sintezo umetnega govora. Na koncu sem rezultate dela primerjal še z enostavnim subjektivnim testom. V uvodu povzemam zgodovino sinteze umetnega govora in sestavo govora samega. Nato na kratko opisujem osnovni model sintetizatorja, ki se je uporabljal pri sintezi ocenjevanih sistemov, in teorijo prikritih Markovovih modelov ter njihovo uporabo pri sintezi umetnega govora. V drugem poglavju predstavljam tako subjektivne kot objektivne metode ocenjevanja govora, s poudarkom na predstavitvi metode DTW in ocenjanja z Mel-kepstrovo razdaljo. V poglavju z rezultati je opis ocenjevanja različnih sistemov za sintezo umetnega govora. Prikazujem ustreznost Mel-kepstrove metode, podajam primerjavo različnih nastavitev metode DTW in ugotavljam, katera od teh je najprimernejša. Zatem predstavljam rezultate subjektivnega testa, s katerim sem preverjal korektnost objektivnih metod. V zaključku razlagam rezultate in predstavljam primernost uporabe takih metod za ocenjevanje umetnega govora ter tudi morebitno nadaljnje delo na tem področju.

• Language: Slovenian
• Keywords: objektivno ocenjevanje, prikriti Markovi modeli, DTW, Mel-kepstrova razdalja
• Work type: Undergraduate thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2016
• PID: 20.500.12556/RUL-85822
• Publication date in RUL: 27.09.2016

Secondary language

• Language: English
• Title: Objective evaluation of speech synthesis systems

Abstract

The purpose of this paper is to evaluate the possibility of using objective measures for the purpose of assessing the quality of synthetic speech. An important step in the development of artificial speech synthesizers product quality assessment. At present subjective methods are the best option for evaluation of text-to-speechsystems, even though they are both time and resource demanding. A method, which would be able to quickly and easily check the quality of the machine itself, as well as different ways of calibrating said machine, would be most useful. In this work we explored two methods that could be used to evaluate synthetic speech called the Mel-Cepstrum distance and DTW (dynamic time warping). The test was carried out on four differenttext-to-speech systems, and evaluated using a simple subjectiveevaluation. The introductory chapter summarizes the history, and composition of artificial speech followed by a presentation of hidden Markov models for the purpose of speech synthesis. The third chapter deals with the subject of both subjective and objective assessment methods where emphasis is given to showing how both DTW and Mel-Cepstrum distances are calculated. In the final chapter, we explain how the evaluation process wascarried out and what the results are. Followed by our interpretation of said results, and possible applications and further development.

• Keywords: objective evaluation, hidden Markov models, DTW, Mel- Cepstrum distance