Gerçek zamanlı konuşma sürümlü jest animasyonu

Gesticulation, which includes instinctive or planned hand, arm and head body gestures, is an essential component of face-to-face communication. Gesture and speech co-exist in time with a tight synchrony, and they are planned and shaped by the emotional state and produced together. In our early studies we have developed joint gesture-speech models and proposed algorithms for speech driven gesture animation. These algorithms are mainly based on Viterbi decoders and can not run in realtime. In this thesis we investigate real-time implementation of these algorithms via optimal adjustment of the parameters in the Viterbi algorithm and focus on synthesizing upper body gestures in real-time, directly from speech signals without need for additional input. Our framework generates upper body gesture animations by selecting gesture phrases, which are defined in terms of body motion extracted from motion capture data. The selection is driven by a pre-trained hidden semi-Markov model (HSMM) which uses prosody features extracted from speech. Experimental evaluations are performed to compare realtime and non-realtime speech driven gesture animations using both objective and subjective evaluations. Objective evaluations quantify the similarity between gesture phrases over the frames of the corresponding animations. Subjective evaluations are performed with A/B pair comparison test. The experimental results confirm that our system is able to produce realistic and compelling speech-driven body gestures in real-time.