A Design of Tree-Structured Filter Bank and It's Application to Noise-Robust Speech Front End

摘要

PURPOSE: A voice recognition preprocessor using a TSFB(Tree-Structured Filter Bank) is provided to sample stronger voice characteristics in noisy environment by using a TSFB and controlling the Q of a channel filter and the gain of channel output of the TSFB. CONSTITUTION: A voice recognition preprocessor consists of a pre-emphasis(110) emphasizing the bandwidth of input audio signals; a TSFB(120) including secondary BPFs(Band Pass Filter) and secondary LPFs(Low Pass Filter) in tree structure, for dividing the bandwidth-emphasized audio signals by each frequency channel; an adaptive Q control unit(130) adaptively controlling the Q of the TSFB by considering input signals of peripheral channels; and an adaptive gain control unit(140) adaptively controlling the gain of each channel output. The adaptive Q control of the channel filters improves the frequency contrast ratio of audio signal channels. The adaptive gain control of channel output enhances the time contrast ratio of audio signal frames.