Investigation of CNN-Based Acoustic Modeling for Continuous Hindi Speech Recognition

Recently, Convolutional Neural Network (CNN) gains more popularity over hybrid Deep Neural Network (DNN) and Hidden Markov Model (HMM) based acoustic models. CNN has the ability to deal with speech signals and it makes appropriate choice for the Automatic Speech Recognition (ASR) system. The sparse connectivity, weight sharing, and pooling allow CNN to handle a slight position shift in the frequency domain. This property helps to manage speaker and environment variations. CNN works well for speech recognition, but it was not appropriately examined for the Hindi speech recognition system. The activation functions and optimization techniques play a vital role in CNN to achieve high accuracy. In this work, we investigate the impact of various activation functions and optimization techniques in the Hindi-ASR system. All the experiments were performed on the Hindi speech dataset developed by TIFR, with the help of the Kaldi and Pytorch-Kaldi toolkit. The experiment results show that the ELU activation function with Rmsprop optimization techniques gives the best Word Error Rate (WER) 14.56%.