Probabilistic Correspondence Mapping for Audiovisual Speaker Modeling

In addition to the framework of probabilistic correspondence mapping on audiovisual speaker modeling, we also explore the correspondence problems with different constraints. Frequency domain correspondence between speakers is established via dynamic programming for speaker normalization in speech recognition tasks. The adjacent constraints in frequency domain actually help to stabilize the algorithm, similar to the dynamic time warping techniques. We also explore the correspondence problem given the manifold structure of different pose face images. It turns out that the manifold structure is very useful to build a good correspondence across different subjects. For audiovisual fusion, a new fusion scheme factorizes audio and visual features into correlated and uncorrelated ones. The correlated features are considered to be the correspondence between two modalities.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical and Computer Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Liu, Ming

Contributors dc:contributor

• Thomas Huang

• Engineering, Electronics and Electrical

Language dc:language

eng

Identifier

(MiAaPQ)AAI3301186

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/81062