Audio-visual quality as combination of unimodal qualities: environmental effects on talking heads

Introduction Talking heads provide a multimodal output component for human-computer-interfaces. They consist of facial visual models that are synchronized with speech synthesis modules concerning speech articulation. Due to their reduction to a human head or upper body, articulation is often more clearly visible compared to a full human body due to the possibly bigger display of the head. Therefore, talking heads are especially suited for applications like robust speech understanding and language acquisition. Evaluation is typically concerned with function test to assess the synthesis quality with e.g. metrics like word error rate of human listeners or perceived naturalness (cf. [8]). But as more and more talking heads are used as interfaces for speech-based dialogue systems and are enhanced with facial expressions, the overall quality experienced by the user is in scope.