Title: The Acoustic Emotion Gaussians Model for Music Information Retrieval
---------------------------------------------------------------------------
Ju-Chiang Wang / National Taiwan University


Abstract:
Developing a computational model that comprehends the affective
content of music and organizes a music collection according to emotion
is essential to music information retrieval. Due to the subjectivity
issues of emotion perception in music, recent years have witnessed a
growing number of attempts that model music emotion as parametric and
probabilistic distributions over the so-called valence-arousal (VA)
emotion space. However, the previous attempts typically did not render
a strict probabilistic foundation. In this talk, I will introduce a
novel probabilistic model, called acoustic emotion Gaussians (AEG),
that realizes the generative process of music emotion perception from
acoustic features. Accordingly, AEG is applicable to both
emotion-based music annotation and retrieval. I will provide graphic
illustration of the model learning process to interpret the rationale
and gain more insights to AEG. Moreover, as AEG is parametric, it is
flexible for personalization via model adaptation techniques in an
efficient and dynamic fashion. At the end of the talk, I will present
an interesting framework that unifies the multi-modal emotion
semantics of music, i.e., the categorical (mood tags) and dimensional
(VA values) emotions. Such a framework, which takes advantage of AEG,
can map mood tags into the VA space in an unsupervised and
content-based learning manner, without any training ground truth for
the semantic mapping. It is consequently applicable to automatically
generate a valence-arousal structured mood tag cloud that comprehends
the discrete semantics over the continuous VA space.
 
---
Ju-Chiang Wang currently is a Ph.D. candidate in Department of
Electrical Engineering, National Taiwan University, as well as a
research assistant in Institute of Information Science, Academia
Sinica, Taiwan. He is expected to acquire the Ph.D. degree in January,
2013. His research interests mainly encompass Music Information
Retrieval (MIR), Machine Learning, and Audio Signal Processing. In
addition to a computer science researcher, cultivated in Taiwan, he is
enthusiastic about playing and creating music in the local rock
bands. His profound understanding in music production, musicology,
signal processing and machine learning gives him particular advantages
in the MIR research, which is by nature multidisciplinary. He won the
first prize of Multimedia Grand Challenge in ACM Multimedia Conference
2012, which is the worldwide premier multimedia conference. He also
built a number of top performed systems for several music
classification tasks in MIREX 2009 and 2010.
 
Website: http://www.citi.sinica.edu.tw/pages/asriver/