Comparative study of hybrid spectral subtraction speech enhancement algorithms

Abstract

Speech is one of the fundamental means of communication. Speech is always accompanied by the background noise thus speech enhancement has been a long standing problem in signal processing. Speech enhancement algorithms are important components in many systems where speech plays a vital part, including telephony, hearing aids, voice over IP, and automatic speech recognizers. Speech enhancement is generally concerned with the problem of enhancing the quality of speech signals. The aim in speech enhancement is to improve the intelligibility and quality of an audio recording. Speech enhancement often involves with poor quality and intelligibility recordings as these audio were under circumstances that did not optimized signal-to-noise ratio. Usually the quality of the audio recordings obtained is damaged by many factors such as poor equalization, additive noise, distortion, or excessive vibration. Many studies involving noise reduction of recorded speech have been founded over decades. Researchers have suggested using hybrid of several algorithms for speech enhancement purpose as it works best, however the comparative studies among these hybrid algorithms are not much found in the field. Speech enhancement can help solves many communication problems and the major achievement is when there are many situations where speech enhancement is used in audio forensics to help reveal the events and able to clarify the dialogue in forensic investigations. In this research three existing hybrid speech enhancement algorithms are studied and compared to see which one gives the best result in speech quality performance. The only way to tell is to evaluate the enhanced speech by using Perceptual Evaluation of Speech Quality (PESQ).