Conversational Digital Humans (CDHs) are increasingly deployed across domains such as healthcare, education, and entertainment, owing to their realistic visual appearance, lifelike motion, and smooth interactive capabilities. However, inconsistencies between a CDH’s visual appearance, vocal timbre, body movements, and interactive textual content can significantly impair the user’s audio-visual experience. To systematically evaluate such multimodal inconsistencies, we introduce the first CDH Quality Assessment (CDHQA) Dataset, comprising 254 videos of eight diverse 3D digital humans. The dataset includes 134 high-quality videos and 120 samples exhibiting three types of quality degradations: mismatches between timbre and appearance, mismatches between actions and interactive text, and combined mismatches. Subjective experiments are conducted to assess the perceptual impact of these issues, revealing their substantial effect on user experience. In addition, we benchmark classical objective quality assessment methods on the dataset, which expose the limitations of existing methods in handling CDH-specific quality issues. These findings underscore the urgent need for more robust and multimodal-aware assessment techniques tailored to CDHs.