Speech Modelling and Sparse Representation Classification for Paralinguistic Applications
Julien Epps
NICTA SML SEMINARDATE: 2013-10-31
TIME: 11:15:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
As systems for recognising linguistic content from speech signals have matured, there has been an increasing focus on paralinguistics a" speaker identity and more recently speaker emotion, gender, age, cognitive load and mental health for example. Speech classification systems for these latter applications are frequently based on acoustic modelling approaches whose review forms the prelude to this presentation. The novel application of approaches based on sparse regression is then detailed, with an emphasis on design pragmatics such as dictionary composition and compensation for unwanted variability, and their advantages are demonstrated on a large database. Finally, some current machine learning-oriented challenges for paralinguistic speech classification are discussed.
