The Mongolian Speech Recognition System Based On Depth Of Neural Network

Speech recognition technology, also known as automatic speech recognition (ASR), whose goal is the human voice in the vocabulary content into computer-readable input, such as keys, binary coding or sequence of characters.In recent years, speech recognition technology has been in full swing in a variety of languages in large, access to all areas. For example, the ubiquitous apple siri system. Unfortunately, high-quality voice service corresponding voice services in minority languages is yet to come. Mongolian speech recognition research has important significance for promoting the prosperity and development of scientific and technological progress and the development of the Mongolian language and culture of the Chinese minority voice information processing.At present, there are more and more publicly available open source software about language and speech processing, while most of them just apply to closed vocabulary. But for applications to handle unrestricted speech input vocabulary, even if again big also can not cover all of the vocabulary. The open-source speech recognition tool has been developed by RWTH Aachen University(RWTH ASR, referred to as RASR), can be combined word units to merge into a new word in the vocabulary, so as to identify the foreign words in the recognition process and realize the large vocabulary continuous speech recognition.This paper introduces the theory of speech recognition technology, and the development of large vocabulary speech recognition acoustic model and speech recognition decoder open-source tool developed at the University of Aachen, Germany (RWTH ASR, referred RASR). Complete signal analysis configuration, estimated Gaussian mixture model and voice decision trees, neural networks combined with the depth (Deep Neural Network, DNN) to give an open vocabulary automatic speech recognition (Automatic Speech Recognition, ASR) system. The main work of this paper is to use neural networks to train a large number of voice data, and get the acoustic model. Meanwhile described in detail how to use open source tools developed RASR continuous speech recognition process, focusing on configuration and implementation of training and recognition.