Deep bidirectional neural networks for robust speech recognition under heavy background noise

Abstract

The Deep Bidirectional LSTM (DBLSTM) with recurrent capability using the old conventional HMM modals performing very well and responding and providing a state-of-the-art technology with very good performance on the on various languages and speech database. Even with the current technologies, the outputs in these works relying on recurrent base neural network schemes. On account of explicit target strategies, it turns out to be extremely hard to coordinate the current day accessible discourse acknowledgment techniques due to the current huge jargon discourse acknowledgment frameworks. This research investigates the possibility of using the DBLSTM as speech acoustic speech miniature in which a calibre neural network scheme with HMM that is a hybrid structure. With few experimental findings we observed that the DBLSTM-HMM hybrid gives better and higher performance and provides better results on various NN schemes and all other previous works the previous works. It is also observed that it outclasses both conventional and deep network benchmarks on a most of the known existing algorithms. The improvements which are occurring in word error rate in the deep network algorithms is modest, considering the fact that a significant elevate in slice level precision. We propose an hybrid technique in which DBLSTM and HMM emerge to be exactly matched for tasks where acoustic modeling and speech prosodies predominates. To add more investigation, need to be conducted on large database and also need to see whether this is working on some specific languages.