Deep neural networks for acoustic modeling – a review

Abstract

Acoustic modeling refers to a statistical model that converts the speech signal to a set of phonetics related to each set of feature vectors extracted through pre processing the sound signal. A traditional approach to this problem is Hidden Markov Models (HMM), a probability model that maps each input with a hidden state. Deep neural networks are used for acoustic modeling due to their efficient feature extraction ability. This paper reviews the various forms of neural networks used in combination with HMMs for speech recognition.