Hidden Markov Model training with side information

In this paper, the iterative Expectation-Maximization equations are mathematically derived for Hidden Markov Models (HMM), when there is partial and noisy access to the hidden states. Since the standard HMM is recovered when this partial and noisy access is turned off, our study provides a generalized observation model; and proposes a new model training algorithm within this model. According to the simulation results, our algorithm can improve the performance of the state recognition up to 70% with respect to the “achievable margin”, and also, is robust to different training conditions.