A study of the HMM for speaker-independent isolated word recognition

Author

Neelakantan, V. ; Gowdy, J.

Author_Institution

Dept. of Electr. & Comput. Eng., Clemson Univ., SC, USA

fYear

1990

fDate

1-4 Apr 1990

Firstpage

90

Abstract

Two new additions in the design of hidden Markov models (HMMs) are proposed. First, a scheme to incorporate the geometric arrangement of the codebook vectors in their vector space is designed. This information is included in the HMM parameters via smoothing after the initial training phase. Second, a scheme to remove the imbalance in the output symbol probability matrix is devised. This scheme involves the use of time-reversed discretized utterances and training the model from the last state to the first one in the training phase of the HMM design. The various schemes were tested on a 2-speaker, 20-word vocabulary task. The results are given

Keywords

Markov processes; speech recognition; HMM parameters; codebook vectors; hidden Markov models; isolated word recognition; output symbol probability matrix; speaker independent recognition; speech recognition; time reversed utterances; training; vector space; vocabulary; Databases; Digital filters; Hidden Markov models; Iterative algorithms; Linear predictive coding; Smoothing methods; Speech recognition; Testing; Training data; Vocabulary;

fLanguage

English

Publisher

ieee

Conference_Titel

Southeastcon '90. Proceedings., IEEE

Conference_Location

New Orleans, LA

Type

conf

DOI

10.1109/SECON.1990.117777

Filename

117777