[Colloquium] Wang/MS Presentation/Dec. 1, 2005
Margaret Jaffey
margaret at cs.uchicago.edu
Thu Nov 17 13:59:15 CST 2005
This is an announcement of Siwei Wang's Master's Presentation.
-------------------------
Date: Thursday, December 1, 2005
Time: 3:00 p.m.
Place: Ryerson 251
M.S. Candidate: Siwei Wang
M.S. Paper Title: Improving Tone Recognition with Combined Pitch and
Intensity Modeling
Abstract:
Tone recognition is an essential component of speech recognition for
tonal languages. Tones produced in continuous speech differ from their
ideal underlying representation due to co-articulation with neighboring
tones,as well as other intonational factors. These influences distort
both tone shape and height and lead to confusion in automatic tone
recognition,though human perception is remarkably robust to these
changes in context.
Based on prior work emphasizing prosodic strength and use of intensity
for tone recognition in devoiced and noisy speech, we develop tone
recognition techniques that augment standard pitch-based approaches
with intensity. Intensity features are employed both for tone region
segmentation to model coarticulatory influences and in the feature
vector for recognition. Using a Sparse Markov Transducer learning
framework on
a corpus of Mandarin broadcast news speech, we achieve a 10-12% absolute
improvement in accuracy over baselines using Hopfield network and SVM
models. Our three-phase segmentation approach outperforms a
simple time-based segmentation by 10%, and classification with intensity
features consistently outperforms pitch-only classification by 15-20%.
We conclude with plans for future work invovling different languages and
learning models
Advisor: Prof. Gina Levow
A draft copy of Siwei Wang's MS Paper will be available soon in Ry 161A.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 161A) (773) 702-6011
The University of Chicago http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
More information about the Colloquium
mailing list