Improved Front-End Analysis in the Army System: Linear Transformations of SRUbank
Martin J. Russell, Darryn Lowe, M.D. Bedworth, K.M. Ponting · Defense Technical Information Center (DTIC) · 1990
Front-end acoustic analysis in early versions of the Airborne Reconnaissance Mission (ARM) continuous speech recognition system was based on the SRUbank filterbank analyser. In its default configuration, this is a conventional, high-resolution filterbank analyser with 27 critical band filters spanning the range 0 to 10 kHz and producing 100 frames per second. This memorandum reports experiments which show that recognition accuracy is improved by applying a suitable dimension-reducing linear transformation to the output of SRUbank. Experiments were conducted using several linear transformations of SRUbank, including 8, 12 and 16 cosine coefficients plus mean channel amplitude, 8, 12 and 16 cosine coefficients plus mean channel amplitude plus difference between corresponding elements of the feature vector at + or - 20 milliseconds, and 8 and 16 principal components. Great Britain.