Spectral refinement with adaptive window-size selection for voicing detection and fundamental frequency estimation

Spectral refinement (SR) offers a computationally in-expensive means of generating a refined (higher resolution) signal spectrum by linearly combining the spectra of shorter, contiguous signal segments. The benefit of this method has previously been demonstrated on the problem of fundamental frequency (F0) estimation in speech processing – specifically for the improved estimation of very low F0. One drawback of SR is, however, the poorer detection of voicing onsets due to the Heisenberg-Gabor limit on time and frequency resolution. This may also lead to degraded performance in noisy conditions. Transitioning between long- and short-time windows for the spectral analysis may offer a good trade-off in these situations. This contribution presents a method to adaptively switch between short- and long-time windows (and, correspondingly, between the short-term and the refined spectrum) for voicing detection and F0 estimation. The improvements in voicing detection and F0 estimation due to this adaptive switching is conclusively demonstrated on audio signals in clean and corrupted conditions.