Output details
15 - General Engineering
University of Edinburgh (joint submission with Heriot-Watt University)
Parametric modelling for single-channel blind dereverberation of speech from a moving speaker
This invited paper shows, for the first time, that time-varying acoustic impulse response (AIR) parameters can be parsimoniously modeled by smoothly-varying basis functions. Using practically measured AIRs, it led to the development of sequential methods necessary for real-time implementation and joint enhancement/localisation techniques investigated in EPSRC/DSTL EP/H012699/1 (£117k). This research, stemming from EP/D051207/1, also presents batch-mode algorithms for speech enhancement from a moving talker by statistical modelling of time-varying AIRs. The technique described will become increasingly important as speech recognition on mobile devices becomes ubiquitous in adverse everyday environments. An invited book chapter resulted (DoI:10.1007/978-1-84996-056-4_8, p.naylor@imperial.ac.uk).