A E Kotze <kotzeae@...>
We have a question relating to obtaining accurate formant values from a script.
If we open a sound in praat, and use "edit" to determine formant values, the spectrogram we see is "normal", i.e. it has a lot of detail, viz. clear striations etc. If we use F1, F2, F3 here we get "good" (acceptable) values for the formants.
However, to obtain values from this window in a script does not seem possible, as a recorded/played back script does not see the Edit window as an object. We understand this, because we know we must access all data and objects from the Praat Object window.
Using the Praat Object Window, if we select Fomants&LPC, LPC, Spectrum (slice) - slice it at point B, and then look at the spectrogram, and use "Get mean" in "Formant Peaks" (setting peaks to 5), we get different formant values using the point B as our slice from the original spectrogram (one accessed from Sound object, using Edit). What is the difference between the ways we are obtaining these values? Also, the image of the spectrogram that we encounter here (in the Object window) is not of a good quality in the sense that it is totally blurred and without detail. Formants also appear as very thin bands. Please indicate if this has something to do with the setting of parameters, or what the reasons for this might be. If so, please suggest parameters. Please also give us some information regarding the setting of bandwidths, i.e. how they are specified.
Winston Anderson & Albert Kotze
Paul Boersma <paul.boersma@...>
AE Kotze wrote:
If we open a sound in praat, and use "edit" to determine formant values,this is not true. You can script an editor from within as well as from a script
that runs from the Objects window. See the scripting manual.
Using the Praat Object Window, if we select Fomants&LPC, LPC, Spectrum (slice) -I assume that you did the following things:
1. Select a Sound object and choose "To LPC". This creates an LPC object, which,
as the manual states, contains linear prediction coefficients that represent the
gross spectral shape of the sound, as a function of time. For speech, this gross
spectral shape is thought to represent the "filter", i.e. the resonances in the vocal tract,
according to the source-filter theory of speech production. The manual contains references
to Fant 1960 and to Markel & Gray 1976. Textbooks on speech analysis explains
this in detail as well.
2. Select the LPC object and choose "To Spectrum (slice)". This creates a sampled
representation of the spectrum defined by the linear prediction coefficients,
i.e. it is a smoothed versions of the spectrum of the sound at the time slice.
If you draw (or Edit) this spectrum for a speech sound, you will see the assumed
filter characteristics of the vocal tract.
Also, the image of the spectrogram that we encounter here (in the Object window)It is the LPC-smoothed spectrum. For speech, it is the filter spectrum,
i.e. the spectrum of the sound without the "source" (glottal) characteristics.
It is based on, say, five formants, and other detail should indeed have gone.
To get the same spectrogram as in the Sound window, you either copy it from
the Sound window to the Objects list, or you select a Sound and choose
"To Spectrogram". Subsequently choosing "To Spectrum (slice)" will give
you the spectrum of the sound at that time slice. See the "Spectrogram" part
of the "Introduction to Praat" in the Help menu.
Institute of Phonetic Sciences, University of Amsterdam
Herengracht 338, 1016CG Amsterdam, The Netherlands