PraatSauce

This set of Praat scripts is designed to facilitate the extraction of spectral measures such as H1-H2, H1-A3, cepstral peak prominence, etc. from acoustic data.

Background and acknowledgments

As the name suggests, these scripts owe a debt to the VoiceSauce suite of Matlab scripts developed by Shue et al. (2011). PraatSauce is not as full-featured as VoiceSauce, but has what some might regard as the advantage of not requiring/being reliant on a Matlab license. Also, thanks to Praat's general Unicode compliance, it permits the use of non-ASCII elements in filenames and in TextGrids.

Most importantly, developing PraatSauce has helped me to appreciate many of the challenges posed by the (semi)automatic analysis of spectral amplitude measurements.

While I referenced the VoiceSauce codebase when developing PraatSauce, be aware that this is not a direct "translation"; PraatSauce makes use of many Praat-internal functions such as To Harmonicity (cc), To PowerCepstrum, etc. As such, identical results cannot be expected, even when using the Praat estimation methods for formants and f0 provided by VoiceSauce. Determining precisely where and why the approaches diverge is the subject of ongoing research; some comparison data are provided for those interested.

The general architecture of PraatSauce is based on a suite of Praat scripts developed by Tim Mills (2010), and as a result PraatSauce is released under the same license as these tools.

Please note that these acknowledgements do not imply any culpability on the part of the named developers. I alone remain responsible for any errors in the present code (but see Reuse and license, below).

I hope PraatSauce will soon be made obsolete by opensauce-python, a Python port of VoiceSauce. However, PraatSauce may still be useful for users who for whatever reason prefer to work in Praat, and/or who find Praat's visualization tools to be useful.

Reuse and license

All scripts are released under the GNU General Public License, version 2.0 or later (the same license as the Praat software). A copy of the latest version of the licence, version 3.0, is provided with the scripts as the file "gpl-3.0.txt". It can also be downloaded from the following URL:

http://www.gnu.org/copyleft/gpl.html

These scripts are provided as-is, without any warranty regarding their suitability or safety for use in any particular application. Users are advised to observe all prudent safety protocols when using these scripts. Neither the author nor the authors of any of the software mentioned above take any responsibility for anything you do or publish based on the results obtained with this software.

Why PraatSauce?

1. Reusability. I learned a lot by implementing the various predecessor scripts to PraatSauce I have created over the years. There are a lot of decisions one needs to make when implementing tools for extraction of voice quality measurements, and it's useful to see how they interact. Putting everything together in a relatively user-friendly fashion seemed like a good way to document my current state of understanding.

2. Reproducibility. If we are going to say in print that it is H1*-A1* that distinguishes voice qualities in Modern Standard Shoggoth, and not some other measure, it would be nice to know that this assertion is robust to at least some degree of variability in the precise measurement technique. It is not currently clear to me that this is the case.

3. Convenience. Matlab is powerful and has excellent signal processing capabilities, but it is not so nice when your site license expires while you are in the field and you can't renew it remotely because it requries an on-campus log-on with an admin key that you don't have. Lots of people use Praat for lots of things so it is often handy to have a Praat-based tool for various tasks. In addition, Matlab doesn't seem to allow Unicode/non-Western characters in TextGrids, anywhere, even if you're not looking at that tier. I have thousands of TextGrids with non-Western characters in them (orthography, usually) and while it's possible to create VoiceSauce-able verisons of those TextGrids, it's a pain in the neck.

4. Curiosity. VoiceSauce includes Praat-based methods for estimating most parameters [1]. How big of a difference can/does implementation make? In theory we should get identical results from identical parameter settings. In practice there seem to be some differences. Not necessarily a problem, but perhaps useful to know.

   [1] Although perhaps crucially, the method of determining harmonic amplitudes is always handled in the same way, which is never Praat-based, regardless of where the formant and bandwidth estimates come from.

What PraatSauce doesn't do

• Many of the fancier things implemented in VoiceSauce: epochs, Sun's SHR methods, etc.
• Alternative methods of estimating F0 (STRAIGHT, Snack)
• Pitch-synchronous calculation of harmonics (yet)
• Correct H2K or H5K for formant frequencies and bandwidths (because this is based on cepstral peak rather than F0 window). Note that VoiceSauce doesn't correct for H5K either, though it does correct H2K.
• Integrate EGG measurements

Requirements

These scripts have been tested with Praat 6.0.36 on macOS Sierra 10.12.6 and Windows 7 Enterprise (SP1). YMMV.

Please note that version 6.0.30 or higher of Praat is required to make use of the new numeric vector object type in the scripts.

Included files

Presently, PraatSauce consists of 8 files, found in the src directory:

• praatSauce.praat, the master script
• shellSauce.praat, a version of the master script with all parameters in a single form, for easy shell scripting
• pitchTracking.praat, which measures f0
• formantMeasures.praat, which measures F1-F3
• spectralMeasurements.praat, which measures harmonic amplitudes, formant bandwidths, harmonics-to-noise ratios, and cepstral peak amplitudes, returning corrections where appropriate
• getbw_HawksMiller.praat, which computes formant bandwidths according to the method of Hawks and Miller (1995)
• correct_iseli_z.praat, which implements the spectral magnitude corrections of Iseli and Alwan (2007)
• splitstring.praat, a helper script which parses a string based on a user-defined separator

Basic workflow

Open and run the praatSauce.praat script. There are several pages of parameters that need to be specified.

Page 1: Directory and files

Input directory and results file

• inputdir is the directory containing your (segmented) audio files.

• textgriddir is the directory containing your TextGrid files. These are expected to have the same name as your audio files, but with the extention .TextGrid. Note that textgriddir can be the same as inputdir.

• outputdir is the directory where the output files will be stored.

• outputfile is the name of the output file. The default is spectral_measures.txt.

session

If you want to start from a specific token (e.g. because the script was interrupted due to some kind of error and you don't want to start all over from the beginning), you can change the value of startToken. You will need to watch the info window to see the token number, or you can figure it out from the Strings list. If you do this, make sure to rename the existing version of spectral_measures.txt as it will be silently overwritten.

Page 2: Intervals and timepoints

channel

If you only want to produce analysis measures for a single channel of a stereo file, enter the channel number here. This is useful if e.g. you have (mono) audio on channel 1 and EGG or another co-registered signal on channel 2. Otherwise, enter 0 to process the stereo Sound file.

interval tier

The tier variable is the tier of interest. Note that every labelled interval on this tier will be analyzed. This means if you have intervals for both e.g. stop closure and rime, you will get spectral measurements for both intervals.

skip these labels

If there are interval labels you don't want PraatSauce to extract measurements for, enter them here as a well-formed regex. ^$ is the empty string (an interval without a label). ^\s$ is a whitespace character (an interval that you meant to leave blank, but has an accidental, invisible space in it). Separate multiple labels with a pipe |. The default regex skips empty intervals, blank intervals, and intervals containing r. See the Praat manual for further details.

point tier

If you have events recorded on a point tier, enter the tier number here. Otherwise leave as 0 (the default).

point tier labels

If point_tier != 0, enter the labels of the events marked on the point tier, separated by spaces. One header column will be added for each label. All other labels will be ignored.

separator

Enter a separator used in the filename (e.g. _ or - or . or similar).

manualCheckFrequency

Although PraatSauce is designed to process data in a 'batch' fashion, formant tracking errors can be common. At least some manual checking of the output is recommended.

Note that presently if the user elects to manually check tokens, the correction will be limited to a single window at midpoint of each labelled interval. Obviously, this is not necessarily going to representative of the harmonic structure at other points in the interval.

measure

This variable allows the user to choose to measure either at n equidistant points, or every n milliseconds. In either case, the n in question is determined by the value of following variable.

points

If points is set to e.g. 5 and measure is every n milliseconds, a measurement will be taken every 5 msec. If measure is n equidistant points than this will take 5 evenly spaced measurements, the precise times of which will depend on the length of the interval of interest.

Page 3: Select measures

The second window allows you to choose the spectral measures you are interested in, along with some analysis window properties. Unlike VoiceSauce, PraatSauce currently just returns a fixed set of measurements. There is also an option to resample to 16 KHz, although this is not really necessary as (a) VoiceSauce does it mostly for STRAIGHT and (b) resampling will occur when formants are estimated in any event.

If you select spectralMeasures but leave formantMeasures and/or pitchTracking unchecked, they will be checked for you anyways, since the spectral balance measures require both Pitch and Formant objects. But it is possible to just get f0 and/or F1-F3 and bandwidths, if you would like (but see Page 4, below).

If you have existing Praat .Formant objects in your inputdir, you may find that PraatSauce is faster if you uncheck the formantMeasures box; you will then load the existing .Formant objects and use these for bandwidth estimation, etc. You can select this option on the following page.

The windowLength and maxAnalysisHz options are used by the formant and spectral amplitude measurement procedures. For formant estimation, windowLength specifies the effective duration of the analysis window in seconds, while maxAnalysisHz gives the ceiling of the formant search range; see Praat documentation for details. windowLength is also used in the estimation of spectral amplitudes to create a 'slice' around a timepoint used to create a long-term average spectrum object. windowPosition is not currently used.

Page 4: Pitch tracking options

Here you can:

• use existing Pitch objects if you have them, or explictly ignore and overwrite them even if you do have them (by leaving the box unchecked)
• change the lower and upper limits for estimated f0. You probably want to change these depending on the speaker's pitch range.

Page 5: Formant measurement options

Here you have the options to:

• play the sound each time a file is processed
• change the time step for the spacing of analysis frames in the formant measurement
• set the maximum number of formants estimated and the point of pre-emphasis
• smooth formant tracks using Praat's Smooth... function, and change the reference formant values for a neutral vowel
• save the visual output as an EPS file
• use existing Formant objects if you have them, or explictly ignore and overwrite them even if you do have them (by leaving the box unchecked)
• decide whether to use Praat's estimates of formant bandwidths, or the Hawks and Miller formula which is the VoiceSauce default. Note that if you haven't selected Pitch estimation two pages previously, the script will exit with an error (because Hawks and Miller's formula relies on a pitch esimate).

Page 6: Got all that?

This pause allows you to save the previous parameter selections as a text file for reference. After this, PraatSauce will begin processing all files in inputdir, pausing only if there is an error or if you have set manualCheckFrequency to be something greater than 0.

Output format

PraatSauce assumes that your filenames contain useful information such as subject codes, repetition number, and so forth, separated by a separator. The output file will include a column for each separator. For example, if your input file has the name 12-cab-w_Audio.wav and the separator is -, the first few of the the output file will look like this:

Filename,var1,var2,var3,Label,seg_Start,seg_End,t,t_ms,f0...
12-cab-w_Audio,12,cab,w_Audio,a,0.648166,0.902471,1,0.648166,136.096...
12-cab-w_Audio,12,cab,w_Audio,a,0.648166,0.902471,2,0.649166,136.192...

t is the rank order of the measurement taken in a Filename for a given Label. t_ms is the absolute time in the file when the measurement was taken.

The abbreviations in the header otherwise follow the VoiceSauce conventions. B1, B2, B3 are formant bandwidths; H1u, H1H2u... are uncorrected harmonic amplitudes or differences; and H1c, H1A1c... are corrected harmonic amplitudes or differences. For more details on the implementation of the corrections, see the CORRECTIONS.md file.

A note on smoothing

VoiceSauce has a smoothing parameter and smooths harmonic amplitudes (but not formants or F0) using a (lagged) moving-average window filter [2]. Currently, there is no smoothing option implemented in PraatSauce, as most data analysis software such as R offers a greater range of easily accessible smoothing options that are fiddly to implement in Praat.

In terms of comparing between VoiceSauce and PraatSauce output: one thing to note is that VoiceSauce appears to first smooths all amplitude measures (H1u, H1c...), uses these smoothed vectors to compute the difference measures (H1H2u, H1H2c...) and then smooths the resulting differences again.

[2] Actually, F0 is smoothed, but using a different procedure.

Calling from the command line

If you would like to script praatSauce from the shell, use shellSauce.praat, which collapses all argument windows into a single window. (If all arguments are included in a single form, which is what is needed for Praat to process all command line arguments when called from the shell due to the single form requirement, the resulting form window will be too big for most screens (since resizing the window is not possible).

For details of how to use Praat from the command line on your operating system, see the Praat manual (currently section 6.9 in Praat version 6.0.37).

References

Hawks, J.W., and J. D. Miller. (1995). A formant bandwidth estimation procedure for vowel synthesis. Journal of the Acoustical Society of America 97(2), 1343-1344.

Iseli, M., Y.-L Shue, and A. Alwan. (2007). Age, sex, and vowel dependencies of acoustic measures related to the voice source. Journal of the Acoustical Society of America 121(4), 2283-2295.

Mills, T. (2010). SpectralTilt-0.0.5 [computer program]. See https://sites.google.com/a/ualberta.ca/timothy-mills/downloads.

Shue, Y.-L., P. Keating, C. Vicenik, and K. Yu. (2011). VoiceSauce: a program for voice analysis. Proceedings of the 17th International Congress of Phonetic Sciences, 1846–1849.

Last modified: Munich, 7 December 2021