Pre-processing speech signals to extract individual variations - normalising amplitude and vector time patterns using logarithmic scale and memory circuit for standardised speech patterns

摘要

A speech signal pre-processing unit for eliminating individual speakers' speech quality patterns and vectors is proposed in order to free the genuine speech characteristics from individual amplitude and amplitude/frequency time patterns before processing in a speech recognition unit. Accordingly, an input terminal for speech signal passes to a speech quality extraction circuit (31). Then the often widely differing amplitude proportions are reduced in a unit (22) by some logarithmic scale and normalized. This is done without destroying the genuine speech criteria and their vector patterns in time. This output is passed to a speech recognition unit (34). A memory circuit (32) holding standardised speech patterns is also applied to the recognition unit (34) and the differences are assessed and evaluated. This corrected speech signal permits analytical treatment in accordance with its objective speech signal contents.