Seminar at Speech, Music and Hearing:
Recording methods for Spoken Dialogue Systems
Opponent: Petra Hagström
AbstractThe aim with this master thesis is to investigate the conditions for generated spontaneous speech and the methods required to produce these kinds of phenomena. To do so one need to
1. Design a scenario which provokes spontaneous dialogues
2. Record the provoked situations with human voices
3. Apply the gained knowledge to generate new and more enhanced synthetic voices
Spontaneous speech consists of much disfluency such as hesitations, false starts and non-lexical utterances. In this project we aim to produce fictional environments where these phenomena can occur and hence record them for further investigation. The purpose of this study is to gain knowledge on how human percept speech and, in addition, apply that knowledge on producing more natural sounding synthesis.
In two of the sessions, the scene set-up and the basic idea behind the game DEAL was adapted. The agents were able to express them-selves freely; however, they were limited to the instructions given by the wizard. The third session was performed in a more laboratory resembling environment. About 12 segments in total were chosen for the listening test that was performed by 16 agents. The result from the test showed that all three syntheses were very good in delivering that an interruption has occurred and the agents were often able to identify the right type of disfluency. However, not all voices were perceived for being natural.
15:15 - 17:00
Tuesday May 26, 2009
The seminar is held in Fantum.
| Show complete seminar list