Loading ...
Sorry, an error occurred while loading the content.

Re: Has anyone seen this Mod to Fldigi source for CW

Expand Messages
  • mni12
    I posted my test audio file in here:
    Message 1 of 36 , May 16, 2012
    • 0 Attachment
      I posted my test audio file in here:
      http://f1.grp.yahoofs.com/v1/0IW0TxWxQ8W-Izb-y-YjGhZbMCVUqRBCak_gvlbB6vDWqeNSd_cYzKtHnuAew80_Xrhg3lAelsrqvjkCb1-QAmv2s0I/ditdah.wav

      It contains THE QUICK BROWN FOX JUMPS OVER THE LAZY DOG - 1234567890
      73 and noise level is high.

      Mauri

      --- In linuxham@yahoogroups.com, "mauri@..." <Mauri@...> wrote:
      >
      > Hi Dave
      >
      > Thanks for the set of test files you provided. I downloaded the files and
      > ran a set of tests.
      > Unfortunately I don't know how to enable Pulse multicast - I am using ALSA
      > Player and running one instance of Fldigi.
      >
      > I did run 9 different tests with the files you provided.
      > Screenshots and explanations on test setup is documented here:
      >
      > http://ag1le.blogspot.com/2012/05/fldigi-more-test-results-on.html
      > I also made some conclusions based on results but it would be good if
      > somebody else can verify these.
      > I hope you can replicate the results above.
      >
      > I am running these tests on Oracle VM Virtualbox v4.1.8 r75467 on a
      > Windows7 64 bit host. I have Ubuntu 10.04 LTS Lynx installed on this VM.
      >
      > What is the best way to send/provide you the wav test file I have been
      > using? It is 3.7 MB in size.
      >
      > 73
      > Mauri AG1LE
      >
      >
      >
      >
      >
      > On May 16, 2012 at 9:28 AM David <w1hkj@...> wrote:
      >
      > > Mauri,
      > >
      > > I pulled Remi's pu/rc source tree and built your experimental CW decoder.
      > It builds fine and does not segfault with audio playback.
      > > To test the decoder I created 9 wav files with s/n ranging from -3 dB to
      > +6 dB. The CW text is a simple CQ sequence. I ran both the experimental
      > and a 3.21.42 version of fldigi simultaneously with a virtual audio feed
      > from a wav player. The simultaneous audio feed was accomplished using
      > Pulse Audio preferences with multicast enabled. This guaranteed that both
      > fldigi instances would be decoding the same audio stream.
      > >
      > > I posted the wav files at http://www.w1hkj.com/developers/
      > >
      > > I tried various matched filter lengths, bandwidths etc, but did not see
      > any noticeable improvement in the experimental decoder. The cpu and memory
      > used by both fldigi versions was nearly identical.
      > >
      > > Would you please repeat my test and let me know what parameters in the
      > experimental decoder gives improved performance. Thanks.
      > >
      > > 73, Dave, W1HKJ
      > >
      > >
      > > --- In linuxham@yahoogroups.com, "mni12" <Mauri@> wrote:
      > > >
      > > >
      > > >
      > > > Hi All
      > > >
      > > > Rene has committed my code in his view, so everyone can access it. The
      > view is pu/rc :
      > > >
      > http://git.berlios.de/cgi-bin/gitweb.cgi?p=fldigi;a=shortlog;h=refs/heads/pu/rc
      > > >
      > <http://git.berlios.de/cgi-bin/gitweb.cgi?p=fldigi;a=shortlog;h=refs/heads/pu/rc>
      > > >
      > > > The software is here
      > http://git.berlios.de/cgi-bin/gitweb.cgi?p=fldigi;a=commit;h=145b7eb1d11a12f54a3d8d6e484c9a3a550c1b0e
      > > >
      > <http://git.berlios.de/cgi-bin/gitweb.cgi?p=fldigi;a=commit;h=145b7eb1d11a12f54a3d8d6e484c9a3a550c1b0e>
      > > >
      > > >
      > > > These two features are still experimental quality and causing Sig 11
      > > > occasionally (especially if FLDIGI configuration: UI/Browser/"Play back
      > history when active channel selected" feature is used).
      > > >
      > > > If anybody has interest in testing these out please have a look at the
      > blog link above and let me know.
      > > > I would appreciate any help making these features more robust and speed
      > optimized.
      > > >
      > > > 73
      > > > Mauri AG1LE
      > > >
      > > > --- In linuxham@yahoogroups.com, "David" <w1hkj@> wrote:
      > > > >
      > > > > When I return from my road trip I will contact Rob to get Mauri's
      > email address. I would like to implement this is fldigi for the alpha team
      > to test.
      > > > >
      > > > > Ed (W3NR), please remind me if I forget.
      > > > >
      > > > > Dave
      > > > >
      > > > > --- In linuxham@yahoogroups.com, Rick Kunath <k9ao@> wrote:
      > > > > >
      > > > > > Has anyone seen this CW mod to the Fldigi source or had the chance
      > to
      > > > > > test the claims?
      > > > > >
      > > > > > I ran across this today... and it looked interesting.
      > > > > >
      > > > > >
      > http://ag1le.blogspot.com/2012/05/fldigi-adding-matched-filter-feature-to.html
      > > > > >
      > > > > > Rick Kunath, k9ao
      > > > > >
      > > > >
      > > >
      > >
      > >
      >
    • Pierfrancesco Caci
      ... Thanks Mauri, all of this is interesting and fascinating. I m sure Dave will put your code to good use. --  Pierfrancesco Caci, ik5pvx
      Message 36 of 36 , May 19, 2012
      • 0 Attachment
        On Sat, May 19, 2012 at 12:51 PM, mni12 <Mauri@...> wrote:
        >
        >
        > I have tested this version mostly against computer generated noisy signals so far as I don't have build environment for Windows 7 OS (I have Flex3000 connected to Windows PC). My main focus was to test if the concept works at all.
        >
        > I started collecting some data using audio recordings from pileups and from stations that have a clear irregular fist. If I could get some help building the code for Windows OS this would go faster.
        >
        > Last year when I developed this concept I created a training dataset  of approximately 40,000 characters recording multiple CW stations in different band conditions etc. I used Octave and Somtoolkit package to run a full training cycle and tested with 20 x 20, 10x10 and 7x 7 sized rectangular SOMs how well the software learns based purely on the dataset. It was quite amazing to see how SOM neural network clustered similar looking morse signals in nearby cells on the grid.  Once the learning was done I did run some classification tests and it worked quite well...only when the signal had a lot of noise it was more difficult to determine the correct class (aka morse character). Note that training dataset contained the shape of the character in 512 long vectors, not the simplified and compressed 7 numbers long duration vectors as in the C++ example.
        >
        > My original idea was to implement the Somtoolkit with C++ and embed this as part of FLDIGI but that would have been a very large project and I don't have so much free time available. Therefore I focused in the first phase to write just the BMU algorithm with C++ and use a fixed SOM table as the codebook.  Since noise seemed to cause problems in classification I started a quest to find a good algorithm to handle that problem. Eventually I found Rob Frohne's KL7NA articles as he had been working on matched filters for quite some time. We exchanged some emails and I got some great insights from Rob that helped a lot.
        >
        > This is a lengthy answer to your question but hopefully gives some background. The current C++ version is a very simplified implementation and it probably has some limitations. I don't know all the limitations as I don't have much free time to put on this project. Therefore I submitted this contribution to FLDIGI community with hope that I could find some other hams to test and improve the code.
        >
        > Dave W1HKJ took my matched filter code and fixed all the problems and improved the efficiency in just few hours. Let's see if somebody can improve the SOM detection in similar fashion.  This is not rocket science but almost as much fun.
        >
        > 73
        > Mauri AG1LE
        >
        >

        Thanks Mauri,
        all of this is interesting and fascinating. I'm sure Dave will put
        your code to good use.


        --
         Pierfrancesco Caci, ik5pvx
      Your message has been successfully submitted and would be delivered to recipients shortly.