Effectiveness of Wavelet Synchrosqueezed Transform for Improved Epoch Estimation from Telephonic Speech Signals Using Zero Frequency Filtering

• Main Authors: Chandni, M., Govind, D.
• Format: Conference Proceeding
• Language: English
• Subjects: Degradation; Epochs; Estimation; Filtering; Instants; Speech enhancement; Time-frequency analysis; Wavelet; Wavelet transforms; WSST; ZFF
• ISSN: 2325-9418
• DOI: 10.1109/INDICON52576.2021.9691652

Epochs are impulse like discontinuities in speech that represent the instants of glottal closure during phonation. The accurate estimation of epochs is essential for epoch synchronous speech processing. There are many algorithms proposed in the literature for the accurate estimation of epochs from speech signals. However majority of those algorithms show degradation during the epoch extraction from telephonic speech signals. The restricted bandwidth (300 Hz-3.4 kHz) of the telephone channel is the reason for degradation in epoch estimation performance. Therefore, the objective of the present work is to propose a wavelet synchrosqueezed transform (WSST) based signal preprocessing stage to improve the performance of zero frequency filtering (ZFF) which is a well known epoch estimation method. Improved time-frequency resolution of WSST reinforce the strength of impulses around the fundamental components to estimate accurate epoch locations. A multiscale product of WSST modified speech and the original speech is computed further enhance the impulse like discontinuities. Effectiveness of the proposed method is confirmed by the epoch estimation performance analysis of phonetically balanced CMU-Arctic database with electroglottogram recordings. The telephone quality in the database is simulated using G.191 software tools.