CROSS-VALIDATION STATE CONTROL IN ACOUSTIC MODEL TRAINING OF AUTOMATIC SPEECH RECOGNITION SYSTEM
St. Petersburg State University; STC Ltd., St. Petersburg; Researcher
M. L. Korenevsky
STC-Innovation Ltd., St. Petersburg; Researcher
K. E. Levin
“Speech Technologies Center”, Ltd; Head of speech recognition department
I. A. Ponomareva
STC Ltd., St. Petersburg; Researcher
N. A. Tomashenko
Laboratory of Computer Science of the University of Le Mans (LIUM), 72085, France; “STC -Innovation” Ltd., 196084, Saint Petersburg ; PhD student; Researcher
Read the full article
Abstract. A technique is presented for optimization of Gaussian mixture models (GMM) size during the training of hidden Markov models (HMM), an essential part of many of the automatic speech recognition systems. Application of the technique increases recognition accuracy by avoiding the over-fitting effect, and reduces significantly computational load of the recognition procedure.
Keywords:
automatic speech recognition, hidden Markov model, cross-validation control, cross-validation criteria
