Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Further Optimisation of a Robust IMELDA Speech Recogniser for Applications with Severely Degraded Speech

Claude Lefebvre (1), Dariusz A. Zwierzyriski (1), David R. Starks (2), Gary Birch (3)

(1) Neil Squire Foundation & Speech Research Centre, National Research Council of Canada, Ottawa, Ontario, Canada
(2) Avionics Division, Canadian Marconi Company, Kanata, Ontario, Canada
(3) Research and Development Division, Neil Squire Foundation, North Vancouver, British Columbia, Canada

Research described in a previous paper [1] demonstrated that high accuracy of recognition of degraded speech is possible to achieve with an IMELDA acoustic representation. The present paper extends these findings and reports on new, incremental improvements to the recognition system. An IMELDA transform is derived for each individual user and it preserves the most salient acoustic features, simultaneously minimising the effects of signal degradation. Increasing recognition accuracy to 99% on speech recorded in a helicopter for the tested population of speakers has been possible through the introduction of a new method of deriving a noise threshold and a modified computation of an IMELDA transform. Problems pertinent to the integration of a prototype recogniser into a helicopter, and preliminary results of in-flight recognition tests are described. Finally, a short section deals with issues involved in computing a transform on a personal computer.

Full Paper

Bibliographic reference.  Lefebvre, Claude / Zwierzyriski, Dariusz A. / Starks, David R. / Birch, Gary (1992): "Further optimisation of a robust IMELDA speech recogniser for applications with severely degraded speech", In ICSLP-1992, 691-694.