The Research Of Rasr Speech Recognition Based On Deep Neural Network

Currently, the public open source software for language and speech processing is more and more available. But the majority of them can only handle the closed vocabulary. For the applications with unlimited speech input, it can’t cover all the words no matter how large the vocabulary is. The open source speech recognition tool has been developed by RWTH Aachen University in Germany(RWTH ASR, referred to as RASR). The word unit in a vocabulary can be merged into a new word by the tool. Therefore, the tool can be used to identify the unknown words and realize the large vocabulary continuous speech recognition.In this paper, in addition to reviewing the theoretical background, the more important thing is to develop the large vocabulary speech recognition system with the RASR toolkit. At the same time, it is necessary to configure the parameters to analyze speech signal, and estimate the Gaussian mixture model. Finally, an open vocabulary automatic speech recognition system based on deep neural network can be realized. The language model is trained by the SRI LM Toolkit. And the SCTK NIST speech recognition scoring toolkit can be used to analyze the results. The focus of the paper is to train the acoustic model with a neural network module, and introduce how to use RASR toolkit to develop a large vocabulary continuous speech recognition system. The most important thing is to configure and realize the steps of training and recognition.