A comparative study of noise reduction techniques for automatic speech recognition systems

Abstract

Automatic Speech Recognition systems are greatly influenced by noise. Noise generated in environment or channel tends to degrade the performance of speech recognition systems. Such unwanted noise signals may alter the main characteristic features of voice signals and corrupt the quality of speech signal and information contained in it. This causes a significant harm to human-computer interactive systems. Noise processing of these signals for speech recognition systems is generally articulated as a digital filtering process in which noisy speech is passed through linear filter to obtain the clean speech estimation. This paper focuses on noise estimation, removal and speech enhancement techniques. In this paper, initial findings support Gamma tone filters instead of conventional Weiner filters and Line Enhancers. Spectral subtraction also showed promising results.