Speech Enhancement Based On Spectral Subtraction Under Non-stationary Environment

IntroductionThe problem of enhancing speech degraded by uncorrelated additive noise, when thenoisy speech alone is available, has recently received much attention. Speech enhancementhas been a classic research topic for its broad application to various speech processingtasks. Many developed techniques are found useful for increasing speech intelligibility,reducing perceptual fatigue and improving speech recognition and speaker identificationsystem. Approaches to retrieve clean speech are plentiful. Among them, spectralsubtraction is one of the most prevailing means under a single channel situation due to itscomputational efficiency.Spectral subtraction presumes that speech and noise signals are uncorrelated.Subtracting a noise spectral estimate from a noisy speech spectrum can therefore retrievethe spectrum of clean speech. The enhanced speech is then reconstructed via an IFFT usingthe modified magnitude spectrum and the original phase spectrum. This type of subtractiveapproach generally results in a reasonably clear quality aside from annoying tonal noise,which is called "music noise". How to reduce the influence of the music noise, especiallyin non-stationary noise environment, is of special importance.1. Noise Spectrum Estimation in Non-stationary EnvironmentA crucial component of a practical speech enhancement system is the estimation ofthe noise power spectrum. A common approach is to average the noisy signal overnon-speech sections. Speech pause detection is either implemented on a frame-by-framebasis. And the detection reliability severely deteriorates for non-stationary environment,week speech components and low input SNR. Additionally, the amount of presumablenon-speech section in the signal may not be sufficient, which restricts the trackingcapability of the noise estimator in case of varying noise spectrum.In this paper, we research a noise estimation approach, namely improved minimacontrolled recursive averaging (IMCRA) is studied, which is the improved form of MCRA.It combines the robustness of minimum tracking with simplicity of the recursive averaging.Improvement the MCRA estimator with regard to the following aspects is conducted:minimum tracking during speech activity, speech presence probability estimation, andderivation of a bias compensation factor. This procedure comprises two iterations ofsmoothing and minimum tracking.Objective and subjective evaluation of IMCRA estimator is performed under...