A SPEECH SIGNAL PROCESSING METHOD BY CONSIDERING FINITE RANGE OF AMPLITUDE FLUCTUATION AND ITS APPLICATION TO ACTUAL ENVIRONMENT

摘要

Speech recognition systems have been applied in various fields due to recent progress of digital signal processing technologies. For speech recognition in actual environment, some countermeasures for surrounding noises are indispensable. Though numerous noise suppression methods for speech signals have been developed up to now, many of those algorithms are originally based on the assumption of Gaussian white noise. However, the real noise signals show complex fluctuation forms with non-Gaussian and non-white properties. Furthermore the speech signal fluctuates within a finite range and its observed data often show amplitude saturation owing to the existence of a finite dynamic range. In this paper, a new method to suppress the undesirable noise in speech signals is proposed in a form appropriate for measured data with finite amplitude range. Though the proposed method requires only a single microphone and doesn't need any priori-information on both noise spectrum and pitch, it works effectively in the presence of noise with high amplitude. More specifically, by introducing a statistical orthogonal expansion expression of the probability distribution based on a Beta distribution defined within a finite fluctuation range, a new noise suppression method is proposed for noisy speech signals. The effectiveness of the proposed method is confirmed experimentally by applying it to real speech signals with white and colored noises.