A Novel N-Average Wavelet Algorithm for a Voice-Based Wheel Chair

Abstract

For the past few decades, speech recognition has been an interesting field of research. Speech recognition has been the best means of communication and it is probably the easy means of communication too. When compared to the other input devices like keyboard and mouse, it is very easy to input and less time-consuming. The speech recognition is achieved through many means of algorithms and architectures. In the field of signal processing and machine learning, the neural network plays an important role. In speech recognition of machines, deep learning algorithms are applied. To train the machines with the consistent speech corpora, classifiers of neural networks like recurrent neural network (RNN), deep recurrent neural network (DRNN), and deep belief network (DBN) are applied. The ultimate target of the deep learning process is to design a machine which could react like a human being, that is, it should learn things like sense organs of humans, remember things like a human brain, learn the action, and recognize things at the times of need. Above all, a language model is very essential for a speech recognition system. The generation of regular grammar, words, and language syntax are formed by a language model and is very helpful to identify the recognition of words easier. The novelty of the work focuses on the generation of a language model for Tamil speech recognition system for making use of a wheelchair for people suffering from permanent or 218temporary or birth disability in order to drive a wheelchair using Tamil voice commands for physically handicapped and aged people. This chapter proposes two approaches for speech recognition. Zero cross rate (ZCR) is used for speech feature extraction in the first approach and secondly, DBN is applied for automatic speech recognition (ASR). For different types of speaker’s identification, DBN is adapted as the effective acoustic model. A novel N-average wavelet algorithm is applied to extract the speech features in the second approach. Adaptive neuro-fuzzy inference system is used to train the noiseless data. The adaptive network-based fuzzy inference system is a type of artificial neural network that is developed based on the inference system which was developed in the 1990s called Takagi–Sugeno fuzzy inference system. By implementing this proposed system more accurate result has been achieved with an accuracy of 99% for Tamil commands and with a maximum elapsed time of about 0.5 s.