Pathological Voice Analysis by David Zhang & Kebin Wu

Author:David Zhang & Kebin Wu
Language: eng
Format: epub
ISBN: 9789813291966
Publisher: Springer Singapore

4.2 Related Works

In this section, many GCI detection algorithms are reviewed and categorized. As to these that are compared with our proposed method in the experiment section, more details are given.

First of all, it is common for numerous GCI detection algorithms to approximate GCIs using various measures in signal-processing field: locations of large values in Linear Prediction Residual (LPR) (Ananthapadmanabha and Yegnanarayana 1979; Drugman and Dutoit 2009; Thomas et al. 2012), lines of maximum amplitudes in wavelet transform (Sturmel et al. 2009; D’Alessandro and Sturmel 2011; Tuan and d’Alessandro 1999), zero crossings after passing the speech to a zero frequency resonator (Murty and Yegnanarayana 2008), and zero crossings of group delay function of LPR (Rao et al. 2007; Naylor et al. 2007; Brookes et al. 2006). Recently, it was claimed that discontinuities estimated under the framework of microcanonical multiscale formalism have relevance to GCIs (Khanagha et al. 2014b). Based on the way to approximate GCI, a majority of GCI detection algorithms can be categorized as follows.

For algorithms based on the LPR, the first step is to compute the LPR from voiced speech. In this stage, speech production is described by the source-filter model and the estimated source is regarded as the LPR. Usually, the filter in this model is simplified as an all-pole type and its coefficients can be estimated by the autoregressive method. With the filter coefficients, the LPR will be obtained by inverse filtering. An ideal LPR signal is an impulse train and the impulse moment within each period coincides with GCI. However, the waveform of LPR for a real speech is much more complicated with many false spikes around the desired peaks (Khanagha et al. 2014b). Thus, multiple approaches are suggested to perform direct or indirect smoothing on the LPR to increase the robustness of GCI estimation. Actually, most algorithms in this category differ only in the way of smoothing. For instance, the Hilbert envelope of LPR was used to find the approximate GCI locations (Rao et al. 2007). The famous DYPSA (Naylor et al. 2007), however, applied the group delay function of LPR to find GCI candidates. In the SEDREAMS algorithm, a mean-based speech was firstly computed to locate a narrow interval for each GCI and then the moment with the highest LPR amplitude inside each narrowed interval was indicated as GCI (Drugman and Dutoit 2009). One variant of LPR is the Integrated Linear Prediction Residual (ILPR). The ILPR signal is obtained by inverse filtering the speech itself (rather than the pre-emphasized speech), whose coefficients are estimated on the pre-emphasized speech. Comparing with the LPR, peaks in the ILPR were demonstrated to be more distinctive (Prathosh et al. 2013). In the YAGA method (Thomas et al. 2012), ILPR was used and a multiscale analysis of ILPR was performed. As suggested by Drugman et al. (2012), smoothing technique is advantageous for LPR-based algorithms. In spite of the direct relationship between GCIs and LPR, LPR is not robust against noise and it often contains peaks of random polarity around CGIs (Murty and Yegnanarayana 2008; Khanagha et al.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.