[BUG] Newly (Today) created LSTM creates aliasing sounding artifacts until fully driven and backed down [VIDEO attached]

[38github]

Describe the bug
When loading the LSTM model the sound is full of what I call "aliasing sounding" artifacts. These however go away when I push the input to max (20dB) and then back down. I have never experienced this before.

To Reproduce
See video.

Expected behavior
No such artifacts

Screenshots
Video
Untitled Project.webm

Desktop (please complete the following information):

• Linux Tumbleweed (x86_64)

Additional context
I used both Neural Amp Modeler LV2 and Ratatouille to demonstrate in the video.

NAM file affected by this
1959hw-crunch-48khz-lstm-nl2hs28.zip

[38github]

Maybe @brummer10 and @mikeoliphant can also help out since I experienced it with their plugins.

[38github]

Maybe this Issue should be in the NeuralAmpModelerCore repository instead? If so I apologise for the inconvenience.

[38github]

I compiled neural-amp-modeler-lv2 with -DLSTM_PREFER_NAM=OFF which I believe makes it use RTNEURAL for LSTMs but it still had issues.

[38github]

Now I have also tried it using Steve's plugin under Windows and it behaves the same way.

[sdatkinson]

Thanks for the bug report.

My intuition is that it's a problem with the initial values that the hidden/cell states are set to.

If I'm right, then if you re-train the model, then it may turn out fine. You should set the RNG seed to something different to be sure.

I don't have time right now to check in more detail, but I'll get it queued up.

[38github]

If I'm right, then if you re-train the model, then it may turn out fine. You should set the RNG seed to something different to be sure.

No problem at all. I will have to read up on RNG seed!

[mikeoliphant]

I wonder if this could potentially be a problem caused by the sigmoid approximation?

@38github To test, you would need to comment out the call to "nam::activations::Activation::enable_fast_tanh();" (which also enables fast sigmoid in LSTM) in whatever plugin you want to test with.

[38github]

I wonder if this could potentially be a problem caused by the sigmoid approximation?

@38github To test, you would need to comment out the call to "nam::activations::Activation::enable_fast_tanh();" (which also enables fast sigmoid in LSTM) in whatever plugin you want to test with.

Thank you Mike for "joining the issue"! Can you give me a hint where I can find and comment out that? Which file? Regards.

[mikeoliphant]

Can you give me a hint where I can find and comment out that? Which file? Regards.

For the lv2 plugin, you need to do it in "/deps/NeuralAudio/NeuralAudio/NAMModel.h":

https://github.com/mikeoliphant/NeuralAudio/blob/84ef71b3189685dbab7a4757f19907dcb9e1d803/NeuralAudio/NAMModel.h#L16

[mikeoliphant]

Make sure you aren't do cmake with "-DLSTM_PREFER_NAM=OFF" - otherwise you'll get the RTNeural version.