SpeakerAICoach: A Multi-Agent Mobile Presenter Training

In this paper, we introduce SpeakerAICoach, a lightweight multimodal framework that evaluates and coaches presentation skills from recorded video. The system decomposes analysis into vision and speech agents that extract verbal and nonverbal cues (filler words, speechclarity, gesturedynamics, gaze, facialaffect, clothing). Moreover, we introduce two state-of-the-art lightweight models for age, gender, and ethnicity estimation, which facilitate speaker demography analysis. An aggregator agent merges per-segment analytics, and an LLM agent generates tailored coaching suggestions. The mobile demo for Android showcases fragment-level feedback, interactive visualizations, and an extensible agent pipeline, which is suitable for on-device clients with server-side inference.

Authors

• Andrey Savchenko Sber AI Lab & ISP RAS Research Center for Trusted Artificial Intelligence
• Anna Slovyagina HSE University
• Irina Bogatyreva HSE University

Keywords

• Multimodal analysis
• presentation coaching
• human-AI interaction