Learning kernel-based HMMs for dynamic sequence synthesis

We propose a statistic model - Kernel based Hidden Markov Model (K-HMM) for dynamic sequence synthesis. From an input sequence, the K-HMM can generate a controlled sequence automatically. A K-HMM is a HMM for which the non-parametric density estimation is used to model the state observation density of the joint input and output distribution. The subtle details of the joint distribution are well kept in our model. We describe the details of learning and synthesizing algorithm of K-HMM. By using a K-HMM, we propose a system that synthesizes a virtual conductor. From a given music sequence, virtual conductor generates a conducting gesture sequence automatically. We demonstrate our virtual conductor by synthesizing extensive animation sequences from input music sequences with different styles and beat patterns.