In this paper, we propose a simple yet very effective feature compensation scheme for two energy-related features, the logarithmic energy (logE) and the zeroth cepstral coefficient (c0), in order to improve their noise robustness. This compensation scheme, named silence feature normalization (SFN), uses the high-pass filtered features as the indicator for speech/non-speech classification, and then the features of non-speech frames are set to be small while those of speech frames are almost kept unchanged. In experiments conducted on the Aurora-2 database, SFN achieves a relative error reduction rate of nearly 50% from the baseline processing.
Bibliographic reference. Wang, Chieh-cheng / Pan, Chi-an / Hung, Jeih-weih (2008): "Silence feature normalization for robust speech recognition in additive noise environments", In INTERSPEECH-2008, 1028-1031.