This paper presents an improved spectral sub- traction method forspeech enhancement. A new noise estimation method is derived in whichthe noise is assumed to be white. By using the property that a whitenoise spectrum is flat, high fre- quency components of a noisy speechspectrum are averaged and the standard deviation of the noise isestimated. This operation is performed in the analysis segment, thusthe spectral subtraction method combined with the new noiseestimation method does not need non-speech segments and as a resultcan adapt to non- stationary noise conditions. The effectiveness ofthe proposed spectral subtraction method is confirmed by experiments.
展开▼