Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.584321
Title: Exploiting the bimodality of speech in the cocktail party problem
Author: Aubrey, Andrew James
Awarding Body: Cardiff University
Current Institution: Cardiff University
Date of Award: 2008
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
The cocktail party problem is one of following a conversation in a crowded room where there are many competing sound sources, such as the voices of other speakers or music. To address this problem using computers, digital signal processing solutions commonly use blind source separation (BSS) which aims to separate all the original sources (voices) from the mixture simultaneously. Traditionally, BSS methods have relied on information derived from the mixture of sources to separate the mixture into its constituent elements. However, the human auditory system is well adapted to handle the cocktail party scenario, using both auditory and visual information to follow (or hold) a conversation in a such an environment. This thesis focuses on using visual information of the speakers in a cocktail party like scenario to aid in improving the performance of BSS. There are several useful applications of such technology, for example: a pre-processing step for a speech recognition system, teleconferencing or security surveillance. The visual information used in this thesis is derived from the speaker's mouth region, as it is the most visible component of speech production. Initial research presented in this thesis considers a joint statistical model of audio and visual features, which is used to assist in control ling the convergence behaviour of a BSS algorithm. The results of using the statistical models are compared to using the raw audio information alone and it is shown that the inclusion of visual information greatly improves its convergence behaviour. Further research focuses on using the speaker's mouth region to identify periods of time when the speaker is silent through the development of a visual voice activity detector (V-VAD) (i.e. voice activity detection using visual information alone). This information can be used in many different ways to simplify the BSS process. To this end, two novel V-VADs were developed and tested within a BSS framework, which result in significantly improved intelligibility of the separated source associated with the V-VAD output. Thus the research presented in this thesis confirms the viability of using visual information to improve solutions to the cocktail party problem.
Supervisor: Not available Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.584321  DOI: Not available
Share: