More specifically, accuracy was assessed by comparing software-generated measurements of synthesized vowels to the input values used to synthesize them. Assessment of accuracy and comparability of acoustic measurements included comparison of acoustic measures including fundamental frequency (F0), first through fourth formant frequencies (F1-F4), and first through fourth formant bandwidths (B1-B4). The purpose of this study was twofold: (1) to make quantitative assessment of the accuracy and comparability of data generated by four AASPs that are most commonly reported in the speech literature and (2) to describe the major features of those four AASPs. Users probably cannot do much to overcome the third and fourth errors rather, these errors should be taken as forewarnings of the limitations of analysis. A parabolic interpolation is particularly problematic when formants are in close proximity. The fourth error relates to the 3-point parabolic interpolation that compensates for the coarse spectrum. The problem is most serious when the roots are close together. The third error is an exclusive reliance on root solving in the LPC algorithm. Vallabha and Tuller (2004) describe a heuristic that can be used to determine the optimal filter order for either a corpus of vowels or a single vowel. There are two common guidelines to set the filter order: (a) set it to the number of formants expected plus two, or (b) set it to sampling frequency in kHz. In general applications, many users adjust the filter order to the speaker only. Users of analysis software should be aware of appropriate adjustments for filter order, taking into consideration characteristics of both the speaker and the speech sample to be analyzed. The second error is choice of an incorrect order for the LP filter. This error is particularly important for the speech of women or young children, who often have an F0 higher than 250 Hz. The first is quantization of the signal owing to the fundamental frequency, which results in an error estimated to be about 10% of F0. Vallabha and Tuller (2002) discuss four sources of error in LPC analysis that are relevant to this study, especially because LPC data are commonly used to generate formant tracks and to populate a data table. LPC analysis has been particularly powerful and convenient because it generates numeric data for formant frequencies and bandwidths that can be displayed in patterns such as formant tracking. Speech AASPs generally afford the capability for FFT and LPC analysis of speech. To our knowledge, there has not been a systematic comparison across AASPs for the measurement of formant frequencies and bandwidths, despite the general interest in these entities for research on typical and atypical speech. In one previous study that compared analysis systems for the measurement of vowel formant frequencies, Woehrling and Mareuil (2007), reported that there were some "substantial differences" in the values of F1 obtained from Praat and Snap. The studies in the latter group raise a concern that values generated by different systems are not always comparable and that care should be taken in managing and interpreting data from these systems. Second, a few studies examined the accuracy and/or reliability of measures of voice, such as the perturbation measures of jitter and shimmer ( Bielamowicz, Kreiman, Gerratt, Dauer & Berke, 1996 Deliyski, Evans, & Shaw, 2005 Karnell, Hall & Landahl, 1995 Smits, Ceuppens & De Bodt, 2005). First, a small number of studies reported on comparisons of features across systems ( Read, Buder & Kent, 1990, 1992) or described general approaches to signal acquisition and analysis without comparing systems ( Ingram, Bunta & Ingram, 2004 Read, Buder, & Kent, 1990, 1992 Vogel & Maruff, 2008). Previous studies of acoustic analysis software packages (AASP) for speech have been of two major types. Furthermore, there is no assurance that the accumulating data gathered from these different systems can be assumed to be accurate and comparable, for healthy or disordered speech, for males or females, or for children as well as adults. However, neutral and objective evaluations of these systems’ measurements have not been reported, so that potential users have little guidance in selecting a system for their use. Use of these software systems is almost certain to lead to a substantial increase in the application of acoustic measures and to the further development of acoustic databases for speech. Software for digital acoustic analysis of speech offers unprecedented opportunities for the analysis of speech samples for different purposes, including education, clinical practice, and research.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |