Speech enhancement based on smoothing of spectral noise floor

This paper presents robust speech enhancement using noise estimation based on smoothing of spectral noise floor (SNF) for nonstationary noise environments. The spectral gain function is obtained by well-known log-spectral amplitude (LSA) estimation criterion associated with the speech presence uncertainty. The noise estimate is given by averaging actual spectral power values, using a smoothing parameter that depends on smoothing of spectral noise floor. The noise estimator is very simple but achieves a good tracking capability for a nonstationary noise. Its enhanced speech is free of musical tones and reverberation artifacts and sounds very natural compared to methods using other short-time spectrum attenuation techniques. The performance is measured by the segmental signal-to-noise ratio (SNR), the speech/ speaker recognition accuracy and the speaker change detection rate for the audio segmentation using MFCC-features (Melscale Frequency Cepstral Coefficients) in comparison to other single microphone noise reduction methods.