Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
Title GPGPU-Based Parallel MFCC using a Multi-Kernel Overlapping
Authors 윤상혁(SangHyeuk Yoon) ; 박능수(Neungsoo Park)
DOI https://doi.org/10.5370/KIEE.2020.69.9.1378
Page pp.1378-1386
ISSN 1975-8359
Keywords Parallel Processing; GPGPU; CUDA; Audio Processing; Machine Learning
Abstract Recently, machine learning applications using audio data are increasing. MFCC is widely used as a feature extraction technique to utilize audio data. It takes a long execution time to compute MFCC with a large amount of audio data. Also, a fast MFCC computation method is necessary for real-time inference. In this study, a GPU-based parallel MFCC using a multi-kernel overlap is proposed to fast compute MFCC. The proposed GPGPU-based parallel MFCC is 434 times faster than CPU-based MFCC. When processing 800 audio data, the proposed GPU-based MFCC using multi-kernel overlap was 2.87 times faster than the previous GPU-based MFCC. Besides, in case processing a long single streaming audio data, the proposed one achieved 1.3 times speed-up compared with the previous one.