Is there a way to run tkassp analyses in batch mode ?
I want to test different parameter settings for some analyses offered by tkassp, either on the same or on different utterance lists. I don't see anyway to do it through the (very well designed) graphical interface, but maybe there is a way to call tkassp through a command line interface ?
Thanks a lot,
Yes there is.
originally, tkassp was a GUI to the assp command line programs and is now in the slow process of changing to different, library based approach. However, the commandline tools are all still there. They can be found in Emu's bin folder. Alternatively, grab the sources from http://libassp.sourceforge.net. The tabs in tkassp are named after the command line tools. An interesting goodie not part of tkassp is afconvert, a versatil program for signal audio file conversion etc.
Help is obtained by issueing the commands without arguments.
btw: thanks for the design compliment
Great, worked like a charm !
This command line functionality makes the signal processing routines very flexible, please keep it available !
I saw there is also a few Gauss windows available for formant estimation with forest. Why not include them in tkassp ? It seems this kind of windowing should be "the best"…
Glad to hear that you like the command line interfaces. If you are using them you'll have access to 3 Gaussian windows ("GAUSS2_5", "GAUSS3_0" and "GAUSS3_5" for alphas of 2.5, 3.0 and 3.5 repectively). Both for tlkassp and the -W option of the command line interfaces I have made a selection of the 24 window functions totally available. The Gaussian windows are not among this selection because I don't consider them "the best" in terms of the trade-off between the width of the main lobe and the highest side-lobe level (spectral vs. amplitude resolution). The Gaussian window with alpha=3.0 and the minimum 4-term Blackman-Harris (with Nuttal improvement) for example have nearly identical main-lobe specs but the highest side-lobe level for the former is about -71 dB and -98 dB for the latter. Clear to me which of the two is "the better". Cheers, Michel Scheffers
Following this page (https://ccrma.stanford.edu/~jos/sasp/Gaussian_Window_Transform.html), I grasped that the sidelobes that arise when applying a Gaussian window come from the fact that the windowing function has to be truncated.
Maybe the sidelobes could be further attenuated compared to other windowing functions either by extending the size of the window (ie. truncating further away from the center) or applying another windowing function, as suggested in the link above ?
Of course in the framework of libassp, this would probably mean implementing a special case to handle this windowing function and/or require more calculation. Maybe it's not worth the hassle, as other windowing function do the job anyway ?
Hi, Yes, it can be said in general that the sidelobes are caused by a window having a finite length. For the Gaussian windows their level is controlled by the parameter 'alpha' through which you can push them down as far as you like. In fact, if you look at te source code in libassp, you'll see that it is implemented as a parametric window with selectable 'alpha'. The consequence of pushing the sidelobes further down, however, that you need to lenghten your window correspondingly to obtain the desired or neccessary spectral resolution. It makes little sense to push them below the signal-to-noise ratio of the audio recording, which is very difficult to get better than 80 dB. This is why I prefer the windows of the Blackman family which reach such values with a shorter window. But it all depends on the kind of analysis/processing you're performing and on the signal.
Log in to post a comment.