A quantitative comparison of glottal closure instant estimation algorithms on a large variety of singing sounds

GCI Detection from Raw Speech Using a Fully-Convolutional Network

“…In order to obtain the ground truth GCIs, we first pre-processed the EGG signals using 5th-order high and low-pass filters with respective cutoff frequencies of 30 and 500Hz. Then, the ground-truth GCI positions were obtained using peak-picking on the negative peaks of the derivative of the pre-processed EGG signal (as done in [12,21]). The pick-peaking was done using the code provided by the authors of [21], available online 2 (getpeaks function in dataloader.py) to have a comparable ground truth for evaluation.…”

“…Until recently, all approaches used to be based on hand-crafted digital signal processing techniques and heuristics. Thorough reviews of those techniques can be read in [11,12,1], where authors compared their performances on a variety of speech and singing signals. Typically, such methods first compute an intermediate speech representation, such as the linear prediction residual [13], a zero-frequency filtered signal [14], or a mean-based signal [15], which emphasizes the locations of glottal closure instants found at local maxima, impulses or discontinuities.…”

“…the mean f0 value for SEDREAMS [15]), and the quality of their results remains quite dependant on the characteristics of the analyzed speech signal (e.g. pitch and voice quality, speech or singing voice) and the targeted application [11,12]. In particular, it has been noted in [12] that all tested algorithms tend to perform better on speech than on singing.…”

“…pitch and voice quality, speech or singing voice) and the targeted application [11,12]. In particular, it has been noted in [12] that all tested algorithms tend to perform better on speech than on singing. Moreover, some algorithms like SEDREAMS [15] or DYPSA [13] also detect some GCIs during unvoiced segments and thus rely on further algorithms to filter out GCI candidates in unvoiced parts.…”

Semi-supervised learning of glottal pulse positions in a neural analysis-synthesis framework

“…Although such approaches have been shown to perform reasonably well, they sometimes rely on manual parameter tuning (e. g., the mean f 0 for SEDREAMS) and the quality of their results remains quite dependant on the characteristics of the analysed speech signal (e. g., pitch and voice quality, speech or singing voice) [10]. Moreover, some algorithms like SEDREAMS [9] or DYPSA [6] also detect GCIs during unvoiced segments and thus rely on further algorithms to filter out GCI candidates in unvoiced parts.…”