Home My Page Projects MPTK: The Matching Pursuit ToolKit
Summary Activity Forums Tracker Lists Docs News SCM Files

Forum: help

Monitor Forum | Start New Thread Start New Thread
RE: Win32 exe : slow mpr.exe [ Reply ]
By: Rémi Gribonval on 2007-05-07 18:18
So far, we have put all our efforts in accelerating the decomposition step (mpd), which was by far the most costly one. We may try to accelerate mpr in the future, if it happens to be a real bottelneck ...


RE: Win32 exe : clipping noise [ Reply ]
By: Rémi Gribonval on 2007-05-07 18:14
We have already met the clipping noise issue a few times, and it seems related to the way we rely on libsndfile to write the output files.

When the analysed signal is normalised (its maximum absolute amplitude value is one), it happens that its approximate reconstruction can take values with magnitude exceeding one, which are clipped when written to a WAV file.

We have not yet come up with a proper solution to fix this in our code, but the following quickfix should work : make sure that the absolute values of the signal you want to analyse do not exceed, say, 0.75, e.g. by renormalisation in Matlab. If does not fix the clipping problem, please let us know.

RE: Win32 exe [ Reply ]
By: Benjamin Roy on 2007-05-03 12:57
I hope this informations will help.



RE: Win32 exe [ Reply ]
By: Benjamin Roy on 2007-05-03 12:55

Most of the MPTK package is fairly stable now. However, we haven't reached version 1.0 yet. So this kind of problem is still possible. Please could you send me the dictionnary file and the signal you used to obtain this result?

Sometime, some problems occurs during the decomposition of a signal using particular dictionary, with the concerned files I could reproduce the problem and try to find its root in order to fix it.

You can make changes in the dictionary structure file (.xml). The rules to write dictionary file are described in the user manual page 7. The dictionary structure file defines the type and the parameter of the atoms used to decompose the signal, here is a example of dictionary structure file that will generate a decomposition of the signal using only Gabor Atoms:

<?xml version="1.0" encoding="ISO-8859-1"?>
<block type="gabor">
<par type="windowLen">128</par>
<par type="windowShift">32</par>
<par type="fftSize">128</par>
<window type="gauss" opt="0"></window>
<block type="gabor">
<par type="windowLen">256</par>
<par type="windowShift">64</par>
<par type="fftSize">256</par>
<window type="gauss" opt="0"></window>
<block type="gabor">
<par type="windowLen">512</par>
<par type="windowShift">128</par>
<par type="fftSize">512</par>
<window type="gauss" opt="0"></window>
<block type="gabor">
<par type="windowLen">1024</par>
<par type="windowShift">256</par>
<par type="fftSize">1024</par>
<window type="gauss" opt="0"></window>

The reconstruction time and the the clipping noise in the reconstructed file are certainly due to this problem during the decomposition of the signal.

Win32 exe [ Reply ]
By: Hassan Lahdili on 2007-05-02 17:37

I'm interested in trying out your tool. I downloaded the windows
version (MPTK 0.5.3-beta for windows) and followed the instructions in the
users' manual.
I was able to run the executable file (mpd.exe) to decompose audio
signals sampled at 44.1kHz but I think something is missing because I got
a kind of clipping noise in the reconstructed file even though I set up
the SNR to 30 dB. One of the messages I got during the process is a
warning regarding an increase in the residual signal's energy (only in
some iterations). I think I'm not using the right dictionnary file
(*.xml). Can you please send me the one you are using for this kind of
applications (decomposing audio signals @44.1kHz)?

Can I make changes to the xml file?

Also, when I used mpr.exe to reconstruct the audio signal (~14 sec
long at 44100 Hz), it took around 55 sec to genrate the audio file. Why
did it take so long?

Your help is very appreciated. Thanks!