A weighted projection measure for robust speech recognition
Beth A. Carlson, Mark A. Clements · 2002
Results of experiments involving a class of low-complexity projection measures which demonstrably improve recognition performance in the presence of background noise are described. The projection measure is used in a continuous density hidden Markov model (HMM) recognition system. The cepstral representation and a perceptually based melcepstral representation for the speech are investigated. The feature vector representation is augmented to include a set of time-differential (delta) parameters, which improved recognition accuracy an average of 10-23%. Of the two representations, the melcepstral showed the greatest improvement. At a signal-to-noise ratio (SNR) of 10 dB, the projection measure resulted in improvement over the standard weighted Euclidean distance in recognition accuracy from 21.8% to 80% for the melcepstral representation and from 35.9% to 70.6% for the cepstral representation.>