SPEAKER DEPENDENT CONTINUO- US SPEECH RECOGNITION OF HINDI LANGUAGE

Abstract

Speech recognition has been a fascinating and interesting topic for researchers for many years. Speech is a natural communication medium with machines especially in the environment where keyboard input is awkward or impossible. Automatic Speech Recognition (ASR) has made great progress for European languages. During the last years great progress in the field has been made, mainly due to many years of research and the availability of high performance systems and algorithms. Speech recognition is a process that converts an acoustic signal to a set of words. In India almost three-fourth of the population lives in rural areas and most of this population is unfamiliar with computers and English. Speaker dependent continuous speech recognition of Hindi language will enable people to interact with computers in their own language and without the use of a keyboard. Speech recognition is used by several different categories of users. People who have difficulty in using their hands to type, professionals, and people with learning disabilities are the main users. A speech recognizer system comprised of two distinct blocks, a Feature Extractor and a Recognizer. The Feature Extractor block uses a Mel-frequency cepstral analysis, which translates the incoming continuous speech in to a feature vectors. Once the features are extracted, they are used for matching with the features of the stored words. The network output is the words to be recognized. For this purpose recognizer block uses neural network. In the course of developing this system, we explored two different ways to use neural networks for audio modeling: prediction and classification. We found that predictive networks yield poor results because of a lack of discrimination, but classification networks gave excellent results. Finally, this thesis reports how we optimized the accuracy of our system with many natural techniques, such as expanding the input window size, normalizing the inputs, increasing the number of hidden units, converting the network's output activations to log likelihoods, optimizing the learning rate schedule by automatic search, backpropagating error from word level outputs, and using gender dependent networks.