Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Missing Feature Theory and Probabilistic Estimation of Clean Speech Components for Robust Speech Recognition

Philippe Renevey, Andrzej Drygajlo

Signal Processing Laboratory, Swiss Federal Institute of Technology, Lausanne, Switzerland

In the framework of Hidden Markov Models (HMMs), this paper presents a new approach towards robust speech recognition in adverse conditions. The approach is based on statistical modeling of noise by Gaussian distributions and an estimation of idealized clean speech directly in the probabilistic domain using a statistical spectral subtraction method and missing feature compensation. The missing feature approach in the probabilistic domain allows the speech features masked by noise to be dynamically detected and estimated in probability calculations performed in HMM based recognizers. The method com-bines the advantages of two techniques: the first based on the statistical compensation similar to the parallel model combination and the second one issued from the missing feature theory.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Renevey, Philippe / Drygajlo, Andrzej (1999): "Missing feature theory and probabilistic estimation of clean speech components for robust speech recognition", In EUROSPEECH'99, 2627-2630.