Bidirectional Octonion Long-Short Term Memory Recurrent Neural Networks for Speech Recognition

Abstract

Over the few decades onwards all the researchers got marvelous attention on machine learning due to a lot of applications in different fields like image processing, speech processing, etc. Automatic Speech Recognition (ASR) is an application of speech processing that achieved tremendous results due to the usage of recurrent neural networks (RNN). A unique type of recurrent neural network is long short-term memory (LSTM), to keeps away from the problem of long term dependencies. In this paper, multidimensional octonion algebra is used to process the input entities with multidimensions efficiently, when compared to real-valued models octonion numbers and octonion neural networks solve many tasks with less learning parameters. We propose a new octonion value long-short term memory (OVLSTM) to efficiently represent long-term dependencies among features in speech sequences prediction. The TIMIT dataset taken for experiments for speech recognition and results are compared with QLSTMS and LSTMs, OVLSTM with less learning parameters reaches better results.