A speech feature extraction system with a linear processor array

A speech feature extraction system for speaker-independent continuous phoneme recognition requires high performance in the 50 MIPS range. The vector processes in feature extraction can be mapped into parallel processes on a linear array vhich consists of multiple processing elements (PEs). Each PE takes charge of the processes of each vector element and extracts features by repeating product-and-sum operations and data/ instruction transfers between tvo neighboring PEs, like a systolic array. The system has a hierarchical configuration with one general purpose digital signal processor (DSP) as the upper layer and a linear array as the lower layer. An experimental system with 16 PEs in the Lower layer is shorn to have about 20 times the processing capability of a single-layered system using only one DSP.