Hidden Markov Models with Imbedded Equalization for Noisy Observations

Signal observations or measurements often contain undesirable but inevitable noisy components which make signal identification related tasks difficult to fulfill. One particular example is speech recognition. A speech recognizer designed or trained under clean or low noise conditions in general suffers serious performance setback when deployed in an environment with a different degree of noise contamination. One way to combat the noise problem is to embed noise characteristics in the signal prototypes, to be used as the reference patterns of the signal in the recognizer. Traditionally, this includes estimation of the noise spectrum and training of the noisy reference patterns. Recently in a separate study [1], we have found that a u unconventional use of cepstral projection distortion measures could lead to robust recognition performance without the need of explicit noise characterization or noisy signal prototypes. This means a recognizer can be designed and trained in a fixed, clean environment and deployed in a unknown noisy environment without specific noise training which sometimes cannot be accomplished due to several reasons.