SIGNAL ENHANCEMENT USING FREQUENCY DOMAIN BINAURAL MODEL FOR HUMANOID ROBOT

摘要

As well known, a "cock-tail party effect" is a binaural effect to emphasize a degraded signal by surrounding noise. Psychoacoustical models of binaural effect have been studied. This paper proposes a frequency domain binaural model using FFT base filter-bank to reduce the computational load. This model is based on both interaural level difference(ILD) and interaural time difference(ITD) of binaural inputs, and weighting balances of ILD and ITD are tuned depending on frequency. The automatic speeh recognition system equipped with this binaural model can reach over 90% correct recognition when direction of noise source is more than 15 degree apart and when SNR=+5dB. Also intelligibility of enhanced speech is quit hight.