Recent technology provides us with realistic looking virtual characters. Motion capture and elaborate mathematical models supply data for natural looking, controllable facial and bodily animations. With the help of computational linguistics and artificial intelligence, we can automatically assign emotional categories to appropriate stretches of text for a simulation of those social scenarios where verbal communication is important. All this makes virtual characters a valuable tool for creation of versatile stimuli for research on the integration of emotion information from different modalities. We conducted an audio-visual experiment to investigate the differential contributions of emotional speech and facial expressions on emotion identification. We used recorded and synthesized speech as well as dynamic virtual faces, all enhanced for seven emotional categories. The participants were asked to recognize the prevalent emotion of paired faces and audio. Results showed that when the voice was recorded, the vocalized emotion influenced participants' emotion identification more than the facial expression. However, when the voice was synthesized, facial expression influenced participants' emotion identification more than vocalized emotion. Additionally, individuals did worse on identifying either the facial expression or vocalized emotion when the voice was synthesized. Our experimental method can help to determine how to improve synthesized emotional speech.