An improved critical-band transform processor for speech applications

This paper presents a computationally efficient algorithm of critical-band transform (CBT) for approximating the critical-band filtering of human ear. The CBT consists of a constant-bandwidth transform in the lower frequency range and a Brown constant-Q transform in the higher frequency range. A 21-band CBT at a sampling rate of 16 kHz is proposed with better approximation to the human critical-band analysis, and significantly fewer computations than other methods. Simulation results also verified its suitability for performing short-time spectral analysis on speech. A VLSI architecture for the hardware implementation of the 21-band CBT is proposed. In comparison to a 256-point in-place FFT processor, the proposed algorithm and architecture result in a 65.7% reduction in on-chip memory accesses, 56.9% reduction in real multiplications and 15.4% reduction in real additions. The proposed 21-band CBT processor is therefore an improved single-chip solution in low-power speech applications.