Efficient Speaker Independent Isolated Speech Recognition for Tamil Language Using Wavelet Denoising and Hidden Markov Model

Abstract

Current research on Automatic Speech Recognition (ASR) focuses on developing systems that would be much more robust against variability in environment, utterance, speaker and language. In this paper all these major factors are considered to develop a system which works powerfully for recognizing a set of Tamil spoken words from a group of people at different noisy conditions. Developing an ASR system in the presence of noise critically affects the speech quality, intelligibility, and recognition rate of the system. Thus, to make a system robust against different noisy conditions, the most popular speech enhancement techniques such as spectral subtraction, adaptive filters and wavelet denoising are implemented at four SNR dB levels namely −10, −5, 5 and 10 with three types of noise such as white, pink and babble noise. This research work is carried out for developing a speaker independent isolated speech recognition system for Tamil language using Hidden Markov Model (HMM) under the above noise conditions. Better improvements are obtained when the proposed system is combined with speech enhancement preprocessor. Based on the experiments 88, 84 and 96 % of recognition accuracy are obtained from enhanced speech using Nonlinear Spectral Subtraction, RLS adaptive Filter and Wavelet approach respectively.