What is the process for writing my own module (not scriptnode)

ustk

@griffinboy said in What is the process for writing my own module (not scriptnode):

@Christoph-Hart

no that's next. I'm either going to use the Vital or Serum method. Serum's is simpler so I might go with that. Precomputed mip maps using fir filters. Easier to program but it will use a lot of memory... Maybe I'll end up switching later down the line to a realtime FFT based mip map method like vital!

Please pardon my dumbness, but I don't really understand mip maps of FIR especially in this context.
To me, it is useful when the waveform is "as is" and not user modifiable, so it is stored with a limited bandwidth upfront.
But if you draw the WFs realtime, how would you get benefit from a mip maps of FIRs?
In this case, wouldn't dropping a single FIR (or any aniti-aliasing filter) at 20kHz (or below the current Nyquist at SR anyway...) be simpler and still effective?

Got a nice lecture on the topic?

griffinboy

@ustk

To start off, putting a single filter at 20k (or Nyquist) on the drawn waveform will not solve the aliasing, because the signal has already been sampled, and therefore the aliasing has already happened. If we oversample the signal by a large amount and then do what you said, yes, this method will work. However now you have a sampler + filter running at 8x oversampling using a ton of cpu. And this can't be precomputed (filtered once only), because if the user plays a higher note, harmonics will pass nyquist again and cause aliasing.

Therefore one solution is to precompute anti aliased versions of the sample at different pitches, removing high frequency content that would alias for specific note ranges. This is called mip mapping. We are essentially decimating the signal (removing high frequency details) for pitches that we play higher up.

I only mentioned fir filters because you can use them also to reconstruct the continuous signal using a certain method detailed in the paper titled 'quest for the perfect resampler'.

Another popular method is to use FFT and cancel out the high frequency bins (vital / serum).

ustk

@griffinboy Thanks for the clarification!

So here we're talking about filtering the drawn waveform, but I was simply referring to dropping a filter after the audio buffer just to cut out those high freqs.
Of course, oversampling always helps, but as you said at a CPU cost that is far from being negligible/ideal...

So each time you draw a WF, is you intent to both filter the drawing AND apply a FIR to the buffers?

As for the FFT bins, this something I was thinking about recently.
I understand the cost in terms of latency introduced by an FFT, cleaning the bins, and reverse FFT isn't something we want for a sampler/realtime instrument.
What is your thought on this for an FX, since we can just report whatever latency we have to the host?
There are different ways to avoid/reduce aliasing, such as usual OS, or more complicated anti-derivative calculation. But since I've never encountered (yet?) such a method implementing a "simple" FFT bin reduction, there might probably be a reason I haven't yet thought about...

griffinboy

@ustk

Right sure.
Yes, you could drop a filter on the audio, but you misunderstand how aliasing works. Aliasing reflects downwards, creating low harmonics. It doesn't just create high harmonics. So even lowpass filtering the signal won't remove aliasing.

You NEED to oversample if you want to filter out aliasing. It's not a case of it 'helping' but it's a way of raising the nyquist so that when you apply the filter, there are no alias harmonics already present, having reflected downwards.

Also, filtering the drawn waveform is the same as filtering the buffer. The waveform is the buffer... Samples along the x axis, y value is the value of each sample.

griffinboy

@ustk

So vital works like this:

Precompute the FFT, store the waveform in the frequency domain.
When the user presses a note, clear harmonics that would alias, convert back to time domain (inverse fft) and play waveform. You only need to calculate the inverse fft once every time the pitch changes, or the user plays a new note. You could also use a threshold similar to mipmapping, say, ignore pitch bends until they go over a certain range. So then you don't do inverse fft constantly when stuff like vibrato is happening.

For antiderivative techniques it becomes even more complicated for pitch varying signals you need to crossfade between mipmaps to avoid clicks.

ustk

@griffinboy Thanks for all that!
I do understand how aliasing works, nyquist, reflected harmonics... But I didn't understand how to prevent it in this particular context until you shine your knowledge on me!
I feel a few percents less idiot now

griffinboy

@Christoph-Hart

By the way, while we are on the topic of Wavetable synthesis, what is your process for scanning through a Wavetable?

Are you doing anything to mitigate large jumps / clicks discontinuities when scanning through different frames? Do you generate any in-between frames using interpolation, or is it simply a case of 'snapping' to the next frame and allowing the realtime interpolator to interpolate between the previous and current sample (belonging to the previous and current frame)?

I'm not in the loop when it comes to the 'common' approach. I'm going to do my own analysis but I thought I'd ask!

Orvillain

@griffinboy said in What is the process for writing my own module (not scriptnode):

@ustk

Right sure.
Yes, you could drop a filter on the audio, but you misunderstand how aliasing works. Aliasing reflects downwards, creating low harmonics. It doesn't just create high harmonics. So even lowpass filtering the signal won't remove aliasing.

You NEED to oversample if you want to filter out aliasing. It's not a case of it 'helping' but it's a way of raising the nyquist so that when you apply the filter, there are no alias harmonics already present, having reflected downwards.

Also, filtering the drawn waveform is the same as filtering the buffer. The waveform is the buffer... Samples along the x axis, y value is the value of each sample.

@griffinboy out of interest... what about avoiding aliasing when downsampling?? I had assumed that a solid biquad lowpass would cover this, but maybe not??

griffinboy

@Orvillain

At the moment I'm using custom fir filters and that seems to work. Again, following the design presented in 'quest for the perfect resampler'

HISEnberg

@griffinboy said in What is the process for writing my own module (not scriptnode):

quest for the perfect resampler

Found it online: Quest for the Perfect Resampler. Thanks for the suggestion @griffinboy

Graham Wakefield's book Generating Sound and Organizing Time also does a good job of covering wavetable synthesis and MipMapping in Max MSP's Gen~ environment, for those who are interested in the topic and need a more "digestible" explanation.

HISEnberg

@griffinboy My personal vision for this would be similar to Max MSP which allows you to install external packages, sort of like expansion packs. This would give full credits to the author of the external nodes (including the licensing structure) and allow building custom libraries for specific purposes. I envisage a system like this as HISE develops and grows to include more developers looking to do different things within the HISE framework.

Christoph Hart

@HISEnberg empty room problem, but as soon as there starts to be demand for it I'll think about a good solution.

Orvillain

@Christoph-Hart

Hey Christoph, when trying to compile your script here, I get:

> Create files
> Sorting include dependencies
> Copying third party files
> Compiling dll plugin
Re-saving file: C:\Users\avern\Desktop\blah\DspNetworks\Binaries\AutogeneratedProject.jucer
Finished saving: Visual Studio 2022
Finished saving: Xcode (macOS)
Finished saving: Linux Makefile
Compiling 64bit  blah ...
MSBuild version 17.10.4+10fbfbf2e for .NET Framework

  Main.cpp
  include_hi_dsp_library_01.cpp
  include_hi_dsp_library_02.cpp
  include_hi_tools_01.cpp
  include_hi_tools_02.cpp
  include_hi_tools_03.cpp
  include_juce_audio_formats.cpp
  include_juce_core.cpp
  include_juce_data_structures.cpp
  include_juce_dsp.cpp
  include_juce_events.cpp
  include_juce_graphics.cpp
!C:\Users\avern\Desktop\blah\DspNetworks\ThirdParty\data_node.h(75,4): error C2593: 'operator []' is ambiguous [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]
!C:\Users\avern\Desktop\blah\DspNetworks\ThirdParty\data_node.h(76,4): error C2593: 'operator []' is ambiguous [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]
!C:\Users\avern\Desktop\blah\DspNetworks\ThirdParty\data_node.h(77,4): error C2593: 'operator []' is ambiguous [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]
!C:\Users\avern\Desktop\blah\DspNetworks\ThirdParty\data_node.h(78,4): error C2593: 'operator []' is ambiguous [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]
!C:\Users\avern\Desktop\blah\DspNetworks\ThirdParty\data_node.h(81,4): error C2593: 'operator []' is ambiguous [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]
	C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.40.33807\include\bit(11): warning STL4038: The contents of <bit> are available only with C++20 or later. [C:\Users\avern\Desktop\blah\DspNetworks\Binaries\Builds\VisualStudio2022\blah_DynamicLibrary.vcxproj]

The lines causing this error are:

hise::JSONObject obj;

// write the values into the JSON object
obj["magnitude"] = magnitude;
obj["RMS"] = rms;
obj["length"] = length;
obj["numChannels"] = channels;

// just attach the current knob position so you know it's working.
obj["knobValue"] = value;

I have just update to the latest develop. This SHA:
03c420c1d12f7a4457a8b497a6b78bc49d250e85

Explicitly casting like this did result in the compile working:

			// write the values into the JSON object
			obj[String("magnitude")] = magnitude;
			obj[String("RMS")] = rms;
			obj[String("length")] = length;
			obj[String("numChannels")] = channels;
			

			// just attach the current knob position so you know it's working.
			obj[String("knobValue")] = value;

griffinboy

@Orvillain

Oh yeah yeah, I found faults with the provided script too. I had to do some casting. I think that's correct.

What is the process for writing my own module (not scriptnode)

36

1.7k

11.7k

102.0k