Design of Approximate Complex Constant Multiplier for Fast Fourier Transform in 28nm CMOS

摘要

Mel Frequency Cepstrum Coefficients (MFCC) has been widely applied in voice interaction for feature extraction, where Fast Fourier Transform (FFT) takes up massive power consumption and area in the hardware. In the low-power serial FFT circuit, complex multiplication is used between process data and twiddle factors. To reduce the hardware cost of the complex multiplication, an approximate complex constant multiplier (ACCM) is proposed to replace the twiddle factor multiplication (TFM) module of the original FFT, which utilizes symmetry mapping relation of twiddle factor to decrease the number of constants of twiddle factor (TF). Meanwhile, approximate processing algorithm is used to increase the common part between fixed constants. Applied on a 128-point 16bit serial FFT in TSMC 28nm CMOS, our method reduces the area of TFM module by 27.4% according to the DC synthesis result at TT 25°C. It has no influence on the MFCC output feature, thus providing an effective and low-overhead method to optimize FFT/MFCC.