A short guide
PeakDetection, <peakdet>, is a MatLab script for calculating fundamental frequency (the inverse of glottal cycle length) and glottal open quotient from electroglottographic signals.
The duration of glottal cycles is measured by detecting positive peaks on the derivative of the electroglottographic signal (hereafter DEGG). The inverse of this duration is still called « fundamental frequency » for short, but there is no check on periodicity, as peaks are detected individually. This is different from correlation-based methods (such as DECOM: Henrich et al. 2004).
To avoid absurd results in the case of double closing peaks, a threshold on maximum fundamental frequency is set by the user, and peaks that are so close that the corresponding F0 is above this threshold are considered as belonging in the same ‘peak cluster’.
In light of the great differences in F0 range across speakers and across the experimental tasks that they perform, it did not appear adequate to set the same threshold for all speakers (say, 500 Hz): the user must be allowed to modify the threshold.
It was not found useful to set a lower threshold parallel to this upper threshold. Implausibly low values in the results point to one of the following situations :
- one closing peak has been detected before the onset of voicing or after the offset of voicing, resulting in the detection of a ‘period’ the inverse of which is under 20 Hz. These cases can be corrected by suppressing the first or last period ; this option is offered by the program, at the stage where the user is asked to check the results.
- some closing peaks within a voiced portion of signal have gone undetected because their amplitude is below the threshold. The user must then check on the figure which amplitude threshold is to be chosen for all the peaks to be detected, and set the amplitude threshold accordingly ; this option is offered by the program when the user is asked to confirm the results.
Smoothing the DEGG signal turned out to be useful in the many cases where there is one single opening peak but its amplitude is very small and it tends to be drowned in noise. The <peakdet> programme smoothes the DEGG signal as follows, where <C_SMOO> is the smoothing step and <dSIG> the derivative (two-point derivative) of the EGG signal :
% filling first and last values with original values
for i = 1 : C_SMOO
SMOO_dSIG(i) = dSIG(i);
SMOO_dSIG(length(dSIG) + 1 - i) = dSIG(length(dSIG) + 1 - i);
for i = 1 + C_SMOO : length(dSIG) - C_SMOO
SMOO_dSIG (i) = sum(dSIG(i - C_SMOO:i + C_SMOO)) / (2 * C_SMOO + 1);
This amounts to : a smoothing step of 1 means smoothing 1 point to the left and right, i.e. each point in <SMOO_dSIG> is the average of 3 points in <dSIG> ; 2 means averaging over (2x2+1) = 5 points.
As the programme computes the open quotient results by four methods two of which operate on the unsmoothed signal, it is avisable to choose a smoothing step (of 1) even if the user believes that this smoothing is unnecessary : this assumption can be verified by comparing the results with and without smoothing, which will show a complete fit if the original signal has very little background noise.
Concerning the choice of a smoothing step for noisy signals : a step up to 5 can be chosen ; visual comparison of the DEGG signals before and after smoothing is recommended to verify that this smoothing does not make neighbouring peaks coalesce. It must be remembered that in some cases the DEGG method simply does not apply, and (arguably) should not be forcibly applied : if taken to its limits, smoothing artificially creates a neat hump for opening and one for closing, but these humps fudge up the issue, as they do not correspond to any precise physiological reality anymore : the advantage of the DEGG method is that it is based on an established relationship between the DEGG signal and significant glottal events ; extreme smoothing blurs this relationship.
In a nutshell : a smoothing step of 1 is adequate for high-quality signals (which already appear visually as very smooth), a smoothing step of 2 or 3 increases correct peak detection in relatively noisy signals. Fudging-up of opening peaks was only observed with a smoothing step of 6 or more.
<peakdet> functions semi-automatically : it allows the user to choose which open quotient results can be retained, on the basis of relevant information : the shape of the signals, and the results of four calculation methods for open quotient. The methods are divided into two sets :
- detection of the local minimum on the signal in-between two closure peaks ; this method is applied twice : on the unsmoothed DEGG signal, and on the smoothed DEGG
- analysis of the shape of opening peaks and calculation of a barycentre of the detected ‘peaks-within-the-peak’, giving each of the peaks a coefficient proportional to its amplitude. Again, this method is applied twice : on the unsmoothed DEGG signal, and on the smoothed DEGG.
The minima method does not take into account the shape of the peak. The barycentre method is sensitive to every change of the sign of the second derivative of the EGG signal, i.e. to every dent in the signal.
Our experience is that detection of the local minima on the smoothed DEGG signal in-between two closure peaks gives precise and satisfactory results in most cases. The barycentre method can be used as a check : if there is a good fit between its results and those obtained by the minima method, the user can confidently assume that the opening peak stands out clearly, and is not dented ; if its results are different, this means that the peak is dented or double, calling for verification by inspection of the signals, and most likely for elimination of open quotient values. This is exemplified in the figures below. Figure 1 shows open quotient results calculated by the four methods. The open quotient curve is reasonably continuous. It starts from a very high value (above 80%), which is frequent after aspirated consonants ; it then goes down to about 55%. The last value is not plausible because there is a sudden jump ; so it calls for confirmation.
The blue line, corresponding to the values calculated by detecting the minima on the smoothed signal, coincides almost entirely with the green asterisks corresponding to detection on the unsmoothed signal, whereas there is a difference (on the order of 2%) between the minima methods and the barycentre methods, the latter yielding lower open quotient values.
The explanation can be found by inspecting the DEGG signal, presented and commented on in figure 2, which shows a close-up view of one of the opening peaks in the corresponding DEGG signal : there is a dent in the signal, the main peak being followed by another peak of smaller amplitude. An intermediate value is computed by the barycentre-method algorithm ; as the secondary peak comes later, the open quotient is slightly lower. On the basis of these observations, the user may decide
- to retain the values calculated by minima detection, considering that they offer a coherent result and that the main peak is salient enough to be chosen as indicative of the timing of glottal opening
- to retain the values calculated by barycentre, which reflect the asymmetrical shape of the peak
- or to exclude all values, considering that the peak is not unique, strictly speaking, which disallows the calculation of a single value of « time of glottal opening » on which to base an open quotient calculation.
The final choice will depend on the objective of the study, and the degree of precision required to demonstrate the hypotheses at issue. If, for instance, the hypothesis was the absence of glottal constriction on the rhyme type illustrated by this example, the experimental uncertainty on the order of 2% would have no importance : all the measurements coincide to indicate that open quotient is near its average value (this observation must of course be refined by proposing speaker-specific reference values) and that there is no tendency to glottal constriction. It would therefore be advisable to include these results despite the slight asymmetry in the shape of the opening peak. On the other hand, if the focus is on differences in open quotient across vowels (‘intrinsic open quotient’), the expectation is that these differences will be small, and an experimental uncertainty on the order of 2% may not be acceptable.
The choice also depends on the quality and abundance of the EGG data. If there are sufficiently numerous cases where there is an excellent fit among methods, it seems reasonable to isolate all the doubtful cases, treating them as a separate set or excluding them altogether.
This part of the work is therefore of great importance, and the user should be as explicit as possible concerning the criteria used.
One example in which open quotient calculation is problematic is shown in figure 3.
The dispersion of the values obtained by the barycentre method hints at imprecise or multiple opening peaks. None of the methods yields a continuous curve of open quotient. It is necessary to exclude all the open quotient values on the second half of the item. On the first half of the item, the results obtained by the minima method applied to the smoothed DEGG signal are not absurd, and are in some degree continuous. It is therefore useful to have a look at the shape of the corresponding DEGG signal : see figure 4, which represents one of the first periods. The peak is strongly dented ; the dents are detected as so many peaks, which results in displacing the barycentre of peaks in a way that varies strongly from one syllable to the next. On the whole, however, there is no clear doubling of the opening peak (which would mean the presence of two peaks of comparable amplitude). This offers an argument for retaining (for the first periods) the open quotient values calculated by the minima method ; the interval over which these results can be retained approximately covers periods 3 to 14. The values for the first two periods, and for the periods after period 15, positively have to be excluded, as there is no salient opening peak. (The choice that we made when treating these data consisted in excluding all the open quotient values, in view of the general variability of the results. If the user chose to retain values 3 to 14, it should be reported that the results are doubtful.)
Figure 5 below exemplifies a frequent situation : the overall fit across methods is good ; the results from the methods applied on the unsmoothed signal contain a few out-of-range values, which are corrected in the results using the smoothed signal, indicating that the quality of the EGG recording is acceptable but not ideal.
Examination of the signal shows that detection of opening peaks is not problematic, though the fine detail of the DEGG signal is not smooth ; this is a case where it could be useful to increase the smoothing step.
In this case, the values in blue (calculation on the smoothed signal : detection of the minima) were chosen, only excluding the value on period 9 because it is higher than its neighbours and the corresponding peak in the signal is not clear.
Figures 7 and 8 show an example in which the barycentre method yields a continuous open quotient curve despite peak doubling. The curve in blue is saw-like : its lack of smoothness indicates that the measurement was not successful, probably due to the fact that the opening peak is doubled or imprecise. The values obtained by barycentre method (black squares) show slightly greater continuity.
Examination of the DEGG signal (figure 8) shows that this is a case of peak doubling. In this instance, the barycentre method appears as an acceptable compromise in case one chooses to provide an approximation of the open quotient despite the absence of a clear, single opening peak. Here again, the final choice (excluding all open quotient values or not) depends on the degree of precision required.
figure 9 (next page) illustrates the excellent fit across methods that is
obtained when the opening peaks stand out clearly, and the corresponding DEGG
signal. In such a case, our choice is to select the method by detection of
minima on the smoothed signal (results in blue) ; the choice of this
method as a default method of sorts is reflected in the choice to plot it as a
stars-plus-line curve, the better to make its continuity (or lack of such)
There are comments within the text of the main program, <peakdet>, as well as in all the functions.
In its present version, <peakdet> requires a .wav file as its input. If there is more than one channel, <peakdet> expects the EGG signal to be in the second channel (right channel of stereo files). If your files are in another format, you can either convert them, or modify <peakdet> using another function such as <allread> for loading the data.
The beginning and end of the portions of the signal to be treated should be provided either (i) as a .txt file containing the beginning and end of each item (in milliseconds), or (ii) as a .txt file produced from a Regions List of the software SoundForge (©SonicFoundry). Here are some details:
(i) The input can be a .txt file containing the beginning and end of each item (in milliseconds) on one single line, separated by spaces or tabs, e.g.
If there are more than 2 columns the software retains the last 2, e.g.
1 2963 3471
2 7662 8428
3 11853 12561
4 16594 17031
5 20863 21466
will be read in the same way as the lines above. This allows you to give the items a number for identification.
(ii) The input can be a .txt file produced from a Regions List of the software SoundForge®. The reason why we did a conversion module specifically for the latter, and not for other formats of annotation, is because we used SoundForge to create the annotation to our EGG files, and needed to be able to make the conversion easily. After doing the annotation in SoundForge, copy it to the Clipboard then to a .txt file.
This file is read and converted by the function <beginend> called by <peakdet>.
When you run <peakdet>, it first asks you to set a few parameters and indicate the path to the .wav file and the .txt file. Then it analyzes the intervals indicated in the .txt file, one by one, asking you to validate the results syllable by syllable ; you can correct the results manually in several ways, suppressing the first or last period, correcting F0 values manually, and suppressing open quotient values that are out of range or correspond to passages where the open quotient cannot safely be calculated. The result is placed in a matrix called <data>, containing one sheet per item analyzed, and padded with zeros. Have a look at the <peakdet> program to see how each sheet is organized ; basically : F0 is in the 3rd column, and the open quotient values retained are in the 10th column.
If you have any questions or suggestions please contact email@example.com
Henrich, N.; d'Alessandro, C.; Castellengo, M.; Doval, B.: On the use of the derivative of electroglottographic signals for characterization of non-pathological voice phonation. Journal of the Acoustical Society of America 115(3): 1321-1332 (2004).