Hardware implementation of a noise robust speech recognition system using RSF/DRA technique

摘要

In this report, we describe a novel method focusing on time variation and dynamic range of speech feature parameters and introduce a hardware based speech recognition system using this technique. RSF (Running Spectrum Filtering) applies the time variation of feature parameters to FIR filtering. DRA (Dynamic Range Adjustment) normalizes the maximum amplitudes of the waveforms. The combination of two methods improves noise robust recognition performance compared with conventional methods. However, the realization of the proposed method in software processing is quite difficult because recognition processing starts after speech analysis processing. Therefore the hardware system accelerates recognition processing remarkably by using parallel processing, fixed-point representation and data scaling.