Most of recent e-learning applications are predominantly based on textual and graphical metaphors to communicate the learning material in its user interfaces. Previous research demonstrated that incorporating both visual and auditory sensory channels in the interaction with e-learning interfaces could enhance the usability and users’ learning performance. Also, the presence of humanoid virtual lecturers in e-learning interfaces was found to be attractive for e-learners. This paper describes an empirical evaluation of three different modes for the employment of speaking avatars as virtual pedagogical agents in a multimodal-based e-learning. This study aimed at comparing the usability aspects (efficiency, effectiveness and user satisfaction) as well as learning performance of three different e-learning platforms recruited to present three common lessons about class diagram notation. The first platform involved talking head of facially expressive avatar while the second platform used a full-body animated one. However, the third tested e-learning platform was based on talking heads of tow facially expressive virtual lectures; male and female who shared the presentation of the learning material. In total, 48 users participated in the experiment to test the three platforms in a within-subject design. Results of the experiment showed that the inclusion of speaking virtual lecturer with full body gestures was the most efficient in terms of question answering time, most effective in terms of correctly answered questions, and the most satisfactory as opposed to the other tow investigated e-learning platforms. Experimental results also revealed that using facially expressive avatars either singularly or coupled with female one scored similar levels of usability and learning performance