Improved hidden Markov models for speech recognition through neural network learning