Light-Gated Recurrent Unit Based Acoustic Modeling for Improved Hindi ASR

Speech Recognition field gets more advantage of deep learning in the last few years. The substantial improvement was reported by Recurrent Neural Network (RNN), which can model long term temporal context. The traditional RNN’s have vanishing gradient problem. This problem was recently tackled by Long Short Term Memory (LSTM) with more promising results. However, the computational requirement of the Bi-directional LSTM model was quite high due to the use of three gating units. Later-on, this computational complexity was reduced by Gated Recurrent Unit (GRU) by using two gating mechanism instead of three. Despite these remarkable achievements, still, there is a need for improvement to beat human accuracy. In this series, we present a novel neural network architecture by combining the recently proposed Light-Gated Recurrent Unit (Li-GRU), Multi-layer perceptron (MLP), and softmax module. In this proposed architecture, MLP was used in the first and last layers as DNNs are good enough in feature mapping and class-based discrimination. In between DNN layers, we stacked Li-GRU layers to take advantage of RNN architecture. This architecture was tested on the Hindi Speech dataset, and noticeable improvement was recorded.