Key points are not available for this paper at this time.
Since wavelet decomposition of signals provides more flexible time-frequency resolutions, it can be utilized as a feature set for speech recognition. The authors explore the possibility of using wavelet decomposition for speech recognition. In particular, they investigate a modified octave structured 5-level filter bank and the HMM (hidden Markov model) is used as a recognizer. We present an analysis of various wavelet filters for speech recognition and compare the results with the conventional features that include LPC and mel-cepstrums.
Kim et al. (Thu,) studied this question.