The next step is to specify exactly how Luscinia is to carry out the comparison. The first two columns of the “Parameter settings” window allow you to vary this.
Dynamic time warping is a dynamic algorithm that provides an overall distance measure for the difference between two elements. The algorithm can be based on any acoustic parameter measured over the duration of the element, resulting in a vector of measurements*. To include a given parameter in the analysis, it is necessary to enter a value in its text box greater than 0.
The weighting values (if greater than 0) represent the relative weightings of the parameters in the analysis. i.e. if one parameter is weighted at 1, and another at 0.5, then variation in the former will carry twice the weight as variation in the latter.
Obviously, variation in frequency measured in Hz is not directly comparable to variation in time in ms. How the different parameters are standardized turns out to be critical to the performance of the algorithm. The method employed by Luscinia - standardization by the joint standard deviation of the two elements being compared - appears to work more effectively than any other obvious method. You can find out more about this here.
On the right-hand side of the window are several other options that alter the outcome of the dtw analysis:
"Compression factor": this determines to what degree Luscinia compresses the raw measures from the database. Compression is carried out according to a logarithmic function: longer elements are compressed more than shorter elements. A value of 1 removes all compression. Values of around 1.2 appear to give good results. More about the compression algorithm can be found here.
"SD ratio" determines the degree to which elements are normalized by the overall standard deviation of the whole dataset (SD Ratio=0) or by the joint standard deviation of each pair of elements (SD Ratio=1). More on this can be found here.
"Offset removal" determines the degree to which differences at one point in the DTW algorithm carry over to the next point. More about this algorithm can be found here. Briefly, what this means is the degree to which overall differences in a parameter like frequency contribute to the distance between two elements. If this parameter is set to 0, overall differences between two signals are weighted weakly, and the overall "form" of the signal are weighted more heavily. If the parameter is set to 1, this is reversed.
"Cost for stitching syllables" is related to "Stitch and compare syllables". One problem I encountered with Luscinia was that it was sometimes difficult to segment syllables into elements in a repeatable way. To counter this, if you select "Stitch and compare syllables", Luscinia will stitch together syllables (after they have been compressed, by the way), and compare the resulting syllables as if they were individual elements. When the computer generates syllable distances later in the process, it will pick the lowest distance score from the ones generated by comparing constituent elements individually, and by comparing the stitched together elements.
You may wish to impose a cost on stitching; that is where the "Cost for stitching syllables" comes in. The distance scores generated by stitching are multiplied by this factor before comparison with those generated by comparing elements separately. In other words, if "Cost for stitching syllables" is set to 1, there is NO cost for stitching.
Obviously, stitching requires that the algorithm takes more time to carry out: in fact the comparison will take about twice as long.
"Weight by relative amplitude" In the basic DTW algorithm,
* You may notice that "Gap between elements" is included as a parameter - but this is obviously not a measurement over the whole element. Luscinia creates a vector, and fills it with the "Gap after element" measure for the element.