Speech segmentation and recognition using adaptive linear prediction algorithm

In this report a speech recognition system implemented with a variable rate segmentation scheme in a point-wise adaptive-lattice-linear-prediction algorithm is investigated. The segmentation procedure developed consists of two unique stages, one determining a locally converged subsegment, and the other, the segment with unique filter state-spectral difference beyond a preselected threshold. The recognition result, using a 40-word alpha-digits vocabulary obtained from three males, shows that the recursive weighting factor (β) of the residual function and the segmentation threshold (θs) are mutually dependent functions and must be optimized concurrently with respect to the sample speech. For speech sample with a 5-kHz bandwith, the β with magnitude of 0.99, and a compression rate near 40%, are found to provide us with the best recognition score. It is understood that such a weighting factor and compression rate would be necessary for an appropriate coding of both the fast changing and slow moving speech sounds, in particular the alpha-digit vocabulary. More data will be added following this preliminary investigation.