Speaker Independent Recognition on OLLO French Corpus by Using Different Features

The Oldenburg LOgatome speech corpus (OLLO) is specifically designed for evaluating speech recognition methods on variability. The performance of features carried on intrinsic variabilities in speech is meaningful for automatic speech recognition (ASR) system. ZCPA and MFCC were the two main features applied to OLLO French corpus in this paper. We took cepstral mean subtraction (CMS) on MFCC. Dynamic transforms (delta-delta-ZCPA and delta-delta-MFCC) were also adopted. The experiments show that the MFCC outperform the ZCPA in separate style. But ZCPA is more robust between different variabilities. The delta-delta operation of MFCC achieves best recognition in noise-free environment. Moreover, ZCPA could be complementary to MFCC so that one can combine them together especially on soft speaking style.