Whispered Speech Detection Using Glottal Flow-Based Features

Recent studies have reported that the performance of Automatic Speech Recognition (ASR) technologies designed for normal speech notably deteriorates when it is evaluated by whispered speech. Therefore, detection useful in order to attenuate mismatch between training and testing situations. This paper proposes two new Glottal Flow (GF)-based features, namely, GF-based Mel-Frequency Cepstral Coefficient (GF-MFCC) as a magnitude-based feature relative phase (GF-RP) phase-based detection. The main contribution proposed features extract magnitude information obtained GF signal. In GF-MFCC, Mel-frequency cepstral coefficient (MFCC) extraction modified using estimated signal derived from iterative adaptive inverse filtering input replace raw similar way, GF-RP modification (RP) instead production provides lower amplitude glottal source than production, thus, via Discrete Fourier Transformation (DFT) information, which make different hypothesized types our are addition, individual GF-MFCC/GF-RP feature, feature-level score-level combination also further improve performance. combinations this study investigated CHAIN corpus. GF-MFCC outperforms MFCC, while has higher RP. Further improved results MFCC (MFCC&GF-MFCC)/RP GF-RP(RP&GF-RP) compared with either one alone. combined score MFCC&GF-MFCC RP&GF-RP gives best frame-level accuracy 95.01% utterance-level 100%.