Learning strategies in speech recognition
Inge Gavăt, O. Dumitru, Claudia Iancu, Gabriel Costache · 2005
Currently, the most successfil approach in speech recognition is based on hidden Markov models (HMM) as clmstjiers and OH mel-Jkquency cepstral coeficients (MFCC) as fealures describing the speech signal. of course, there are altematives like support vector machines (SYM) and artijcial neural networks (m) m claslfiers and perceptive linear prediction (PLP) coefficients as features. In this paper we will first compare the per$ormmces obtained in a digit recognition task applying SYM and HMM as class$ers and MFCC and PLP coeficients ds features. AIthough the clussic merhod (MFCC+HMM) gives slightly better results, the icse of SYM in speech recognition seems to be attractive. The second, task for comparison is vowel recognition where we found veg> encoiiruging the fact that A?CN acts better than HMM.