Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
This paper presents a newly formulated speech, recognition algorithm for keyword spotting which uses a feature enhancing artificial neural network, a semi-continuous hidden Markov model, and a likelihood ratio test based on optimal detection theory to make decisions regarding possible keywords. The speech recognizer can be used to detect the occurrences of a single word within connected input speech streams in noisefree neutral or Lombard stressed environments. A keyword-dependent neural network  enhances speech, parameters and reduces the probability of false acceptances of non-keywords by adapting its weights and input layer width based on extracted speech characteristics . Using the neural network reduces false acceptances by more than for mono-syllable keywords in a defined keyword spotting application . Enhanced features are submitted to a semi-continuous hidden Markov model which produces a score indicating the presence of the represented keyword. A likelihood ratio test uses functions formed from keyword and non-keyword recognizer training data for detection. Receiver operating characteristics (ROC's) show that the new recognition algorithm can improve keyword spotting performance for neutral and Lombard effect speaking conditions.
Bibliographic reference. Clary, Gregory J. / Hansen, John H. L. (1992): "A novel speech recognizer for keyword spotting", In ICSLP-1992, 13-16.