Modified magnitude spectral subtraction methods for speech enhancement

摘要

In this paper, the method proposed by Boll, noise is estimated from the initial few frames of silence region and is subtracted from noisy speech spectral magnitude, the subsequent proposing steps are residual noise removal and additional signal attenuation to improve the performance. In this research work, we proposed to study the effect of non-stationary noise and use of over subtraction factor and spectral floor factor as proposed by Berouti. The experimental results indicate that variations implemented results in better performance for helicopter noise and all types of noises at negative SNR's, the spectrogram also indicate that this method does the job similar to that of Boll's proposal.