Research And Optimization Of Speech Enhancement Algorithm Based On Short-Time Spectrum Estimation

The production, transmission, and reception of the speech signals, as is thecase with practical production, are subjected to interferences and influence ofvarying degrees, occurring as a result of background noises which hinder thenormal reception of speech signals, with a consequent suppression ofperformance of speech processing system. The attempt to improve the quality ofspeech signals results in our method which consists of using s peech enhancementtechnique to accomplish speech enhancement of noisy signal and improving thespeech definition and intelligibility by extracting the original speech signal aspure as possible from the corrupted speech signal in the receiver.This paper is focused on the two main methods, used in short-time spectralestimation speech enhancement algorithm, namely, the spectral subtraction andminimum mean square error estimation algorithm (MMSE), and features theefforts to improve the two algorithms.The conventional traditional spectral subtraction designed for speechenhancement suffered from a greater disadvantage, the occurrence of obviousmusical noise remaining in enhanced speech signals. The measure taken toeliminate the drawback is a multi-band spectral subtraction. The improvedalgorithms works by dividing the noisy signal into varying bands according tofrequency and preventing them from overlapping each other and then obtainingthe reduction factor of each band by calculating the noise ratio and usingadaptive algorithm, depending on each band of the noisy signal and the noisesignal.The study consists of analyzing the minimum mean square error estimationalgorithm (MMSE) and coupled with conventional algorithm and speech perception characteristics, applying maximum likelihood estimation to the logspectral for the improved algorithm. The key technology consists in reframingeach signal by maximum likelihood estimation and introducing two dynamica llyadjusted parameters and β into the estimation, thus improving the flexibility ofalgorithm.The final simulation of improved algorithms proves the advantagesdemonstrated by the improved algorithm, such as a greater effectiveness, reducedmusical noise occurring in enhanced signals, improved speech definition andintelligibility, increased signal-to-noise ratio, and what is more, a significantreduction in white noise, thanks to the specially improved MMSE algorithm.