Speech-Driven Automatic Facial Expression Synthesis

This paper focuses on the problem of automatically generating speech synchronous facial expressions for 3D talking heads. The proposed system is speaker and language independent. We parameterize speech data with prosody related features and spectral features together with their first and second order derivatives. Then, we classify the seven emotions in the dataset with two different classifiers: Gaussian mixture models (GMMs) and Hidden Markov Models (HMMs). Probability density function of the spectral feature space is modeled with a GMM for each emotion. Temporal patterns of the emotion dependent prosody contours are modeled with an HMM based classifier. We use the Berlin Emotional Speech dataset (EMO-DB) [ 1 ] during the experiments. GMM classifier has the best overall recognition rate 82.85% when cepstral features with delta and acceleration coefficients are used. HMM based classifier has lower recognition rates than the GMM based classifier. However, fusion of the two classifiers has 83.80% recognition rate on the average. Experimental results on automatic facial expression synthesis are encouraging.