Neural networks and neural fuzzy systems for speech applications

Abstract

Speech is the most used means for human beings communication. Among the major constraints for general applications of automatic speech recognition, the presence of unavoidable background noise is of great importance since the speaker cannot be isolated to obtain a clean acquisition of the uttered speech to be processed. The separation of noise and speech with traditional _ltering techniques is hard since respective spectra overlap each other in frequency-domain. One of the successful approaches to remove noise from speech is the adaptive noise canceling which aims to subtract noise from a received signal in an adaptive manner. Several algorithms developed under this approach have shown good performances. Optimization techniques like neural networks and fuzzy inference systems offer the potentiality to deal with encoded data by learning and reasoning as humans do. Neural Networks have a remarkable learning capability such that a desired input-output mapping can be discovered through learning by examples. Their use in an adaptive noise cancellation system, allows the compensation of channel effects. On the other hand, Fuzzy Inference Systems success is mostly due to the fact that fuzzy if-then rules are well-suited for capturing the imprecise nature of human knowledge and reasoning process. The resulting system that combines these two schemes produces robust systems for noise cancellation problem. We report, on this dissertation, an analysis in integrating both neural networks and fuzzy inference systems for noise cancellation from speech signal. We implemented a nonlinear model to cancel noise from the speech signal based on Adaptive Neuro-Fuzzy Inference System and exploiting Widrow's approach. The system is tested in different noisy environments. Performance studies are carried out for qualitative and quantitative evaluations. Time is an important factor when designing a system for a given task. A fast neurofuzzy inference system is proposed for speech noise cancellation in almost real-time computation. Since noise is distorted by a highly nonlinear process before corrupting the speech signal, we proposed some passage dynamic functions that model the environment through which the noise waves undergo before corrupting the speech signal. Moreover, we implemented a neuro- fuzzy system able to identify the noise source before canceling its waves from the speech signal. Classification is an important problem in the pattern recognition _eld. Before being used by the speech recognition system, speech sentences must be preprocessed. Two preprocessing algorithms namely Linear Prediction Coding (LPC), and RASTA (Relative Spectral) techniques are used for extracting features from the speech signals to classify both English stops and noise soures. Time Delay Neural Networks and Recurrent Neural Networks are used as classifiers.