Third European Conference on Speech Communication and Technology

Berlin, Germany
September 22-25, 1993


Robust Speaker-Independent Speech Recognition Using Non-Linear Spectral Subtraction Based IMELDA

Helge B. D. Sorensen (1,2), Uwe Hartmann (2)

(1) Department of Electronics and Electrical Engineering, The Engineering Academy of Denmark, Lyngby, Denmark
(2) Speech Technology Centre, Institute of Electronic Systems, Aalborg University, Aalborg, Denmark

A two-stage noise reduction approach for robust speaker-independent speech recognition is presented. The application is speech recognition in the presence of stationary and non-stationary car noise. The two-stage method calculates noise robust NSS-IMELDA feature vectors, which are used as input to Continuous Hidden Markov Models (CHMM). The new NSS-IMELDA features are based on an integration of two noise robust feature extraction methods - the Non-linear Spectral Subtraction (NSS) method [1] - developed by Lockwood et al. and the Integrated Mel-scale with Linear Discriminant Analysis (IMELDA) method [2] developed by Hunt et al. The NSS-IMELDA features result in better speech recognition in noise than pure an NSS based - or an IMELDA based recognition system. Results on the test database consisting of the test part of the TI-DIGITS database [3] and a non-stationary car noise database gave recognition rates of 98.8% and 98.9% at 0 dB and 15 dB, respectively.

Full Paper

Bibliographic reference.  Sorensen, Helge B. D. / Hartmann, Uwe (1993): "Robust speaker-independent speech recognition using non-linear spectral subtraction based IMELDA", In EUROSPEECH'93, 235-238.