Butterfly framework of LPCC ASIC design for friendly HMI in speaker identification

This work proposes a friendly and affordable speaker identification system using a low-cost linear prediction cepstral coefficients (LPCC) in a application-specific integrated circuit (ASIC) design. To achieve the advantages for human-machine interface (HMI), this system requires the feature performing in a fast responding time and a low cost design on the embedded platform. Consequently, two modes, that is, the butterfly architecture (BA) mode and the multiply-accumulate-tree (MAT) mode, are proposed to be switchable in the LPCC ASIC design. The MAT mode employs the butterfly architecture to reach the lower access time and the smaller controller area compared to the LPCC baseline. The BA mode provides the capability of resource sharing and parallel processing to lessen the computation burden. In contrast to the baseline, the simulation results indicate that the proposed accelerator is capable of decreasing 37% of area cost, 58% of responding time, and an accuracy rate in 90%. Experiments exhibit the entire performance of the proposed work is greater than the baseline, and thereby demonstrates the effectiveness of the proposed work.