Noise Estimation Using High Frequency Regions for Spectral Subtraction

Junpei YAMAUCHI  Tetsuya SHIMAMURA 

Publication
IEICE TRANSACTIONS on Fundamentals of Electronics, Communications and Computer Sciences  Vol.E85-A  No.3  pp.723-727
Publication Date: 2002/03/01
Online ISSN: 
Print ISSN: 0916-8508
Type of Manuscript: LETTER
Category: Digital Signal Processing
Keyword: 
spectral subtraction,  noise estimation,  high-frequency regions,  non-stationary noise condition, 

Full Text: PDF(737KB)

Summary: 
This paper presents an improved spectral subtraction method for speech enhancement. A new noise estimation method is derived in which the noise is assumed to be white. By using the property that a white noise spectrum is flat, high frequency components of a noisy speech spectrum are averaged and the standard deviation of the noise is estimated. This operation is performed in the analysis segment, thus the spectral subtraction method combined with the new noise estimation method does not need non-speech segments and as a result can adapt to non-stationary noise conditions. The effectiveness of the proposed spectral subtraction method is confirmed by experiments.