Evaluating the modified viterbi decoder for long-term audio events monitoring task

Our developed solution of intelligent audio surveillance systems is based on modified Viterbi decoder customized for long-term audio-events monitoring application. The classical Viterbi decoder evaluates the hypothesis after detecting the end of the event (utterance) in the front-end module. The end of the event is usually detected using UBM (Universal Background Model) or relative entropy or energy features with adaptable threshold, etc. Our modified decoder does not require assistance from front-end module and the hypothesis are evaluated upon our proposed algorithm by decoder. This paper describes the evaluation experiment which should approve that the decoder modification will have no significant impact on audio events recognition accuracy.