Rule-Based Speech Morphing for Verification of Emotional Perception Model

摘要

This paper reports rules for morphing speech voice to make it can be perceived as different primitive features, for example, to make it sounds more like "bright" or more like "dark". For the perception of emotional speech, we have proposed a three-layered model, which contains five categories of emotional speech, primitive features, and acoustic features, in previous work. The concept is based on that we assume that human perceive emotion from speech is according to a combination of different primitive features that listeners give to the utterance they hear. Based on experiments and acoustic analysis, we have built the relationships between the three layers by a top-down method and reported that such relationships are significant for perception of emotional speech. In order to verify the relationships, a bottom-up method is adopted. That is to morph (resynthesize) speech voice by composing acoustic features in the bottommost layer to produce speech voice with perception of single or multiple primitive features, which can be further perceived as different categories of emotion. To this end, the rules of morphing speech to make it be perceived as different primitive features should be established first. This paper describes how the rules were established and presents the principles and the strategy of rule creation. An experiment was conducted to evaluate the performance of rules. The results of the experiment indicate the proposed principles and strategy are a direction to create rules of primitive feature. And, the results also confirm the relationships between primitive features and acoustic features built previously are related and effective.