The authors have studied the conventional MEM (maximum entropy method) and developed an MEM algorithm for a precise spectrum analysis method in short-duration frames very little affected by noise. The new algorithm requires about twice as large a word length and ten times as large a processing speed as conventional spectrum analysis algorithms, so it is implemented in the form of a dedicated LSI. This LSI is a DSP-type LSI with a 32-b*16-b multiplier, a 56-b ALU, and a barrel shifter within a chip. A speaker-independent voice recognition system featuring accurate identification of phonemes can be constructed by combining this LSI with a voice-pattern-matching LSI.
展开▼
