Extraction of expression from Japanese speech based on time-frequency and fractal features

The extraction method based on time-frequency and fractal features was proposed to analyze intonations from Japanese speech signal. Two parameters were presented to reveal different feature patterns: Peak spectrum (F max) and Fractal dimension (FD) trajectories. The F max and FD were computed by using short-time Fourier transform (STFT) and Higuchi´s method, respectively. Speech data recorded from 15 Japanese utterances, 4 different ways of expression (accosting, wholehearted, normal, and uninterested). The results showed that the proposed features could extract different intonations statistically in comparison with baseline intonation.