This work presents a novel approach for the design of a predictive model of music that can be used to analyze and generate musical material that is highly context dependent. The system is based on an approach known as n-gram modeling, often used in language processing and speech recognition algorithms, implemented initially upon a framework of Variable-Length Markov Models (VLMMs) and then extended to Variable-Length Hidden Markov Models (VLHMMs). The system brings together various principles like escape probabilities, smoothing schemes and uses multiple representations of the data stream to construct a multiple viewpoints system that enables it to draw complex relationships between the different input n-grams, and use this information to provide a stronger prediction scheme. It is implemented as a MAX/MSP external in C++ and is intended to be a predictive framework that can be used to create generative music systems and educational and compositional tools for music. A formal quantitative evaluation scheme based on entropy of the predictions is used to evaluate the model in sequence prediction tasks on a database of tabla compositions. The results show good model performance for both the VLMM and the VLHMM while highlighting the expensive computational cost of higher-order VLHMMs.