Discriminative training of the pronunciation networks

Author

Korkmazskiy, F. ; Juang, B.H.

Author_Institution

Dialog Syst. Res. Dept., AT&T Bell Labs., Murray Hill, NJ, USA

fYear

1997

fDate

14-17 Dec 1997

Firstpage

223

Lastpage

229

Abstract

Presents a new approach for a pronunciation network construction. The structure of a pronunciation network is determined as a result of the acoustical data decoding procedure that evaluates a list of the N most probable strings of pronunciation units (SPUs), such as phonemes. The importance of each of the decoded strings is characterized by a set of weight coefficients prescribed to phonemes or to some part of the phonemes. The optimality of the weight coefficients is defined in the framework of discriminative training, and the use of the minimum classification error (MCE) criterion allows us to maximize the discrimination between different pronunciation networks

Keywords

decoding; hidden Markov models; learning (artificial intelligence); probability; speech recognition; acoustical data decoding procedure; decoded strings; discriminative training; minimum classification error criterion; most probable pronunciation unit strings; optimal weight coefficients; phonemes; pronunciation network construction; speech recognition; Acoustics; Decoding; Dictionaries; Hidden Markov models; Loudspeakers; Network topology; Radio access networks; Speech recognition; Statistics; Training data;

fLanguage

English

Publisher

ieee

Conference_Titel

Automatic Speech Recognition and Understanding, 1997. Proceedings., 1997 IEEE Workshop on

Conference_Location

Santa Barbara, CA

Print_ISBN

0-7803-3698-4

Type

conf

DOI

10.1109/ASRU.1997.659009

Filename

659009