Jump to content


  • Content Count

  • Joined

  • Last visited


About pkane2001

Recent Profile Visitors

3325 profile views
  1. Thanks, Rando. Is this just with stereo playback or also happens with a single channel?
  2. Spectrogram is just a number of FFTs done on consecutive steps of the waveform. Since each step is already fairly small, the last FFT may contain very little actual data to perform the frequency analysis. It's possible that the very last part of the spectrogram may look very strange compared to the rest. This will vary with the total number of samples and the selected spectrogram settings. It's probably best to ignore the last step, or I can even omit it from the plot. Here's an example where the last 10-15 samples appear to cause a problem in the last step (extreme zoom-in):
  3. A quick follow-up release, v1.0.40 fixes a few regression items introduced in v39: Re-enabled dynamic volume adjustment while playing music from the main window Fixed the wrong color used in Spectrogram when all values are the same (e.g., 0dB) Fixed “Stopped! Matrix dimensions must agree: op1 is …, op2 is …” error Thank you for testing!
  4. As suspected, the plotting library decided to recompute the min and the max values for the color range if the values don't span the whole spectrum (of colors). In effect, the issue would only arise if all the values in the spectrogram are the same (e.g., 0dB). Shouldn't cause any real problems, but I'll definitely fix it.
  5. Thanks for testing, Arpiben! It should be displaying the delta spectrogram as the difference of the two spectrogram values, both in dB. The same two files should result in 0dB. When I subtract the same two files, the color seems to indicate a very large result, but the actual value is 0dB, so the color mapping is off somehow. Let me look into it.
  6. Rookie mistake! I deleted one line I was sure wasn't needed just before doing the last build, and of course, it was necessary! Please uninstall the previous version and re-download and re-install from the same link. I didn't change the version number. Regards, -Paul
  7. Hmm. Very strange. Didn't happen at all in my testing.
  8. Thanks, Arpiben! I may know where this is happening but need to confirm. Does this happen on more than one set of files?
  9. File types currently supported for reading (all should work on Windows 10, a couple may not work on previous versions of Windows): FLAC, WAV, MP3, AIFF, DSF, DFF, AAC, ALAC, OGG/Vorbis. For exporting, only WAV file format is supported in 32- or 64-bit floating point.
  10. Not much going on lately, seems things are pretty quiet and stable. Time to try to break things, again Please try version 1.0.39b. This has some bug fixes and new features: Changes in 1.0.39b Support for processing, playing, and comparing/exporting stereo files Added new selector for filters to filter Ref, Comp, Ref+Comp Added amplitude range selector for Spectrogram plots Improved FIR filters (LP, HP, and Notch) with better out-of-band rejection Improved aliasing behavior resampling operation to reduce artifacts Cleaned up some control UI sizing and painting issues Proper application of all the new FFT Windows introduced in 1.0.38 More details 1. Two channel stereo files can now be loaded, matched, played, exported, and compared using ABX comparator. Because of the double size of files in memory compared to mono, shorter files or a large amount of free memory is highly recommended! 2. You can now select to filter reference and/or comparison files independently using Filter 1, 2, and the Notch filter: You can now click the red X to the right to reset all filter settings to off and to zero out all trim settings in one click. Select one of the R+C filter options to filter both, Reference and Comparison files (that was the automatic behavior in previous versions). 3. Spectrogram amplitude range (dB) can now be adjusted to better highlight a portion of the range, mapping the selected range to the full spectrum of colors: Above, the delta spectrogram is zoomed on the region between -5 and +5 dB, as an example. 4. FIR filters, in combination with the new, better FFT Windows can now produce better out of band rejection results: This is a combination of high pass, low pass and a notch filter just on the Reference file. 4. Much cleaner resampler behavior with regards to aliasing. Both, upsampling and downsampling produces a cleaner result: This is an example of a chirp generated at 96kHz sampling rate, and downsampled to 44.1kHz. Previous versions of DW were using Windows resampler and produced something like this:
  11. Right. Interesting, but again, not earth-shattering. At least in computer models, a recurrent neural network can recognize and process much more complex patterns than a simple feed-forward network. That's because the patterns are computed not just over the weights of individual connections but also over time, with the result of previous process being combined with the input to the next cycle. This allows the same neural net to process lowest level details and highest level abstracted symbols and everything in between. The article makes it sound like the brain is filling in most of the visual details, but the paper is much more discrete in the magnitude of their claims
  12. Just looked, thank you. Makes much more sense than the article summary So what was your take away from this paper?
  13. Don't know. The thing that was interesting to me, right at the beginning of the article: The part about their specific model wasn't as interesting to me, as there was very little detail given. Recurrent neural networks (loops as they call them) are nothing new. They've been studied and modeled many years ago. Not everything is known about them, but it's not news.
  14. Don't believe it was double-blind. Do you have ABX logs to prove it?
  15. Well, yes, the article was about vision and not hearing. But the "layers and layers" of processing that overlays what we see with reconstructed, and sometimes imagined, details is the point I got out of it.
  • Create New...