Recognition of acoustical alarm signals for the profoundly deaf using hidden Markov models

A new acoustical alarm signal recognition scheme for tactile hearing aids using hidden Markov models (HMM´s) is presented. In particular, a maximum likelihood classifier is proposed where the observation probability density function of each alarm class is modelled by a four-state HMM. The performance is evaluated using a database of 205 alarm signals from four typical alarm classes, and is compared with a conventional minimum-distance classifier and with a neural network approach. The results show a superior recognition performance of the HMM-based classifier when compared with the mentioned alternatives. The presented recognition scheme is well suited for real-time implementation due to its low computational costs