EUROSPEECH 2001 Scandinavia
7th European Conference on Speech Communication and Technology

Aalborg, Denmark
September 3-7, 2001


A Multiconditional Robust Front-End Feature Extraction with a Noise Reduction Procedure Based on Improved Spectral Subtraction Algorithm

Bojan Kotnik, Zdravko Kacic, Bogomir Horvat

University of Maribor, Slovenia

In this paper, the procedure for feature vector extraction in multiconditionally noisy environments is presented. Proposed front-end uses time and spectral domain processing for noise reduction as well as feature extraction to create mel-cepstrum parameters and achieves a trade-off between effective noise reduction and low computational load for real-time operations. First, a novel weighting function is used to reduce the rough noise in time domain, and then a spectral subtraction method based on minimum statistics is applied to decrease the effect of additive broadband noise on speech in the spectral domain. At final stage, a feature vector, which consists of 12 mel-cepstrum parameters and the energy, is created. For evaluation of improvement of speech recognition with presented front-end, the "Aurora 2" database together with the HTK recognition toolkit have been chosen. With proposed method an average improvement in performance of 24.75% relative to the current ETSI Aurora standard was achieved.

Full Paper

Bibliographic reference.  Kotnik, Bojan / Kacic, Zdravko / Horvat, Bogomir (2001): "A multiconditional robust front-end feature extraction with a noise reduction procedure based on improved spectral subtraction algorithm", In EUROSPEECH-2001, 197-200.