Hilbert-Huang Mel Frequency Cepstral Coefficient for Speech Stress Recognition System

Abstract

Today, acoustic and spectral characteristics are commonly utilized to determine stress levels, and they have a high degree of accuracy. Mel frequency cepstral coefficients are the most successful spectral characteristic (MFCCs). On MFCC framework, each window was determined using a Fourier transformation based on its resolution. The window's size, on the other hand, causes frequency resolution issues, particularly for under-stressed speech, since the frequency of each speech can change due to emotional conditions. Hence, it is necessary to analyze the frequency spectrum over time and has an adaptive window. To address this issue, we apply Hilbert-Huang transform into MFCC (called HFCC) feature extraction technique for more robust stress speech recognition system. The extracted features are then used as trained data using neural network (NN) to identify the emotional stress of speaker. We used actual speech stress data from SUSAS Database in all experiments. The experimental result shows that HFCC outperforms MFCC and the existing feature extraction techniques.