A method for finding a maximally representative graph structure for the acoustic events associated to linguistic units is presented. The method consists in a three-stage Dynamic Programming technique, applied to sets of acoustic label sequences collected for linguistic units, and obtained by Viterbi alignment of the incoming speech against the states of an Ergodic HMM. These sequences are first aligned in between them, and then collapsed into a reduced trellis. The trellis is then repeatedly explored by a second DP procedure which selects a list of the more probable sequences. Finally, a third algorithm is utilized for dynamically build ah HMM network.
Some experiments on the HMM Structure Inference / Falaschi, Alessandro; P., Pierucci. - STAMPA. - (1990), pp. 1375-1378. (Intervento presentato al convegno Proceedings of Eusipco-90, Fifth European Signal Processing Conference tenutosi a Barcelona nel settembre 1990).
Some experiments on the HMM Structure Inference
FALASCHI, Alessandro;
1990
Abstract
A method for finding a maximally representative graph structure for the acoustic events associated to linguistic units is presented. The method consists in a three-stage Dynamic Programming technique, applied to sets of acoustic label sequences collected for linguistic units, and obtained by Viterbi alignment of the incoming speech against the states of an Ergodic HMM. These sequences are first aligned in between them, and then collapsed into a reduced trellis. The trellis is then repeatedly explored by a second DP procedure which selects a list of the more probable sequences. Finally, a third algorithm is utilized for dynamically build ah HMM network.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
