Minimum based Noise Suppression for Improved Automatic Speech Recognition

J. Fernndez, C. Meyer, and A. Fischer (Germany)


Noise suppression, Spectral Estimation, Automatic SpeechRecognition


In this paper, we propose a simple and efficient noise sup pression algorithm for robust speech recognition, using first order statistics. The basic idea of our approach is to use the minimum of the current signal and its time average in indi vidual frequency bands as an approximation of the noise estimate in subsequent spectral subtraction, thus avoid ing over-subtraction of the noise. Evaluation is carried out on the AURORA-II connected digit strings database. While being computationally very simple, we achieve up to 36% (12%) relative word error rate reduction on the clean (multi) task, respectively, compared to a standard proba bilistic voice activity detection algorithm including non linear spectral subtraction.

Important Links:

Go Back