A remarkable skill that is often taken for granted is the ability of the human cognitive system to distinguish between multiple simultaneously active speakers. In fact, in 1953 a challenge was proposed to the engineering community:
"How do we recognise what one person is saying when others are speaking at the same time? On what logical basis could one design a machine for carrying out such an operation?". Blind source separation (BSS) is one such approach to solve this problem, where BSS is the recovery of original source signals from a given set of mixed observations with minimal a priori information on the environment. There are numerous methods to BSS, from those with a highly statistical basis to those motivated by the human auditory system. In particular, the auditory phenomenon of masking, where components in the perceived speech mixtures with lower energy are suppressed whilst the higher energy components are emphasised, has been of interest to researchers of BSS. This phenomenon has been realised within the time-frequency masking approach to BSS, and this concept of masking for source separation has emerged into its own field of research. This seminar explores the clustering-based approach for estimation of such separation masks. We present a comparison between different clustering algorithms and investigate their suitabilities in a range of simulated and real-world environments. Furthermore, the majority of BSS methods assume that the number of sources is known to the system. As such, we also consider the automatic detection of the number of sources by introducing some novel clustering-based approaches to the problem. Short bio: Ingrid graduated from UWA in 2009 with a Bachelor of Engineering and went on to study a PhD with the Signal and Information Processing Group of UWA. She competed a research internship at Nippon Telephone and Telegraph Corporation in Kyoto, where she worked under the supervision of Dr. Tomohiro Nakatani in the Signal Processing Research Group of the Media Information Laboratory. Ingrid submitted her PhD thesis recently in 2014, where the focus of her thesis was on the clustering-based approaches to blind source separation and source number estimation. She is currently at the Centre for research experience, and she is working in the field of directional analysis for ocean waves. PS* This seminar is free and open to the public & no RSVP required. ****All Welcome****
|
|
© 2001-2010
The University of Western Australia Questions? Mail [email protected] |
