Fourier & Wavelet Methods for Finding Speech Onset Latencies

Abstract

<p>Localization of speech onsets to determine onset latencies is a complicated problem with as many different methods for finding them as there are different areas which use such measurements. A majority of research performed in cognition uses a standard amplitude threshold voice key for estimating the speech onset latencies but a number of studies have shown that this method is incredibly inaccurate and can bias data or produce contradictory results. A number of alternative methods based on modifications to traditional voice-keys have been proposed to deal with the inconsistency although still show a number of deficiencies. Previous work has suggested that switching from the amplitude domain of a signal to the frequency domain a number of the issues present with voice keys can be overcome and when used in conjunction with a number of highly sensitive heuristics highly accurate onset latencies can be produced reliably under ideal conditions. This research is refined and paired with a new user interface to improve the ease of use and increase the adoption rate of this type of analysis. Recent work in the telecommunications industry also suggests that wavelet-based algorithms in conjunction with the Teager Energy Operator (TEO) can accurately detect speech even in the presence of noise. Four wavelet-based methods are investigated and tested; a simple wavelet transform test, and three methods using wavelet-packet transforms in conjunction with the TEO. Although these methods do not perform very well compared to traditional methods a number of potential issues with the implementation are discussed.</p>