Comparison of Acoustic Model Adaptation Methods and Adaptation Database Selection Approaches

Speaker adaptation methods has very important role in modern speech recognition systems as adaptation of acoustic models usually improves accuracy of spech recognition system significantly for specific speaker. This paper evaluates and compares approaches to select data used for adaptation. The appr...

Full description

Saved in:
Bibliographic Details
Published in:Journal of electrical and electronics engineering 2010-01, Vol.3 (1), p.147
Main Authors: Papco, Marek, Juhár, Jozef
Format: Article
Language:English
Subjects:
Accuracy
Actuaries
Data compression
Discriminant analysis
Regression analysis
Voice recognition
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Speaker adaptation methods has very important role in modern speech recognition systems as adaptation of acoustic models usually improves accuracy of spech recognition system significantly for specific speaker. This paper evaluates and compares approaches to select data used for adaptation. The approach of selection of adaptation data with aim to increase the occurence of rarely used phones is compared with randomly selected adaptation data. The SEMI-TIED, HLDA and MLLR approach are used to adapt HMM models with 4,8,16 and 32 Gaussians. The results of tested basic and adapted models show that with MLLR adaptation were achieved the best acuracy improvement. We achieved average improvement of 11,85% for male speaker and 15,78% for female speaker. [PUBLICATION ABSTRACT]
ISSN:1844-6035
2067-2128