Convolutional Grid Long Short-Term Memory Recurrent Neural Network for Automatic Speech Recognition

The Grid Long Short-Term Memory (Grid-LSTM), which is consisted of three steps, i.e., two-dimensional grid splitting, local feature projection, and grid sequence modeling, has been widely used in Automatic Speech Recognition (ASR) tasks, since it has a strong time-frequency modeling ability. However, the network suffers from a serious problem that heavy computing time is always required. It can be found that the reason for this problem is in the last step, two cross-working LSTMs are employed to model time-frequency features in the grid via an analysis of its process. Thus, we try to speed up the Grid-LSTM by using a smaller grid and propose two enhanced Grid-LSTM models, i.e., Convolutional Grid-LSTM (ConvGrid-LSTM) and Multichannel ConvGrid-LSTM (MCConvGrid-LSTM) to reduce the grid size from the two dimensions of the Grid-LSTM respectively. In the frequency axis, we try to do this by using a large frequency stride and further to prevent performance loss by embedding a CNN in the Grid-LSTM. Moreover, in the time axis, we model several adjacent frames by the multichannel processing ability of CNN. Our method achieves 54% relative reduction of training time and 19% relative reduction of Word Error Rate (WER) for a character level End-to-End ASR task.