Synthesis of listener vocalizations : towards interactive speech synthesis

Abstract

Spoken and multi-modal dialogue systems start to use listener vocalizations, such as uh-huh and mm-hm, for natural interaction. Generation of listener vocalizations is one of the major objectives of emotionally colored conversational speech synthesis. Success in this endeavor depends on the answers to three questions: Where to synthesize a listener vocalization? What meaning should be conveyed through the synthesized vocalization? And, how to realize an appropriate listener vocalization with the intended meaning? This thesis addresses the latter question. The investigation starts with proposing a three-stage approach: (i) data collection, (ii) annotation, and (iii) realization. The first stage presents a method to collect natural listener vocalizations from German and British English professional actors in a recording studio. In the second stage, we explore a methodology for annotating listener vocalizations -- meaning and behavior (form) annotation. The third stage proposes a realization strategy that uses unit selection and signal modification techniques to generate appropriate listener vocalizations upon user requests. Finally, we evaluate naturalness and appropriateness of synthesized vocalizations using perception studies. The work is implemented in the open source MARY text-to-speech framework, and it is integrated into the SEMAINE project's Sensitive Artificial Listener (SAL) demonstrator. Dialogsysteme nutzen zunehmend Horer-Vokalisierungen, wie z.B. a-ha oder mm-hm,…