[LAD] Re: Tape emulation - is it worth the effort ?

Hi,

Maybe using an actual tape player and neural-amp-modeler, considering 
the tape player is "an amp" ? 
https://neural-amp-modeler.readthedocs.io/en/latest/tutorials/gui.html

I am not sure about the quality of this idea, but I am confident members 
of linux-audio-dev list are able to tell the good and bad for such an 
approach.

Also, such a plugin could be usefull for sound design.

On 10/17/24 07:35, Fons Adriaensen wrote:
> On Thu, Oct 17, 2024 at 12:27:30AM +0200, [email protected] wrote:
>   
>> Regarding your approach of white box reverse engineering, have you had a
>> look at Jatin Chowdhury's work on the subject?
>> (https://github.com/jatinchowdhury18/AnalogTapeModel).
>> He went quite far simulating the full chain and his plugin has acquired
>> quite a reputation.
> Yes, I know of his work. But reading the paper, I don't think
> his physical model is correct.
>
> Unless I'm missing something, the algorithm he uses amounts to
>
> - upsampling, giving S(t)
> - adding the bias signal B(t)
> - evaluating the hysteresis loop on S(t) + B(t), giving M(t)
> - downsampling, which filters out the HF bias.
>
> But that NOT how things work in reality.
>
> Imagine following a small fragment of tape (corresponding to
> one sample in the upsampled domain), as it moves past the
> recording head and away for it.
>
> The magnetic field of the record head extends well beyond
> the head gap, decreasing in strength as you move away from
> it.
>
> So the small tape fragment will go through to many
> cycles of the bias + signal waveform with decreasing
> amplitude. The resulting magnetisation is not just
> the 'average' of M(t).
>
> For an example, see
> <http://kokkinizita.linuxaudio.org/linuxaudio/downloads/magnetisation.png>
>
> X-axis is time in microseconds, bias frequency is 100 kHz,
> bias amplitude is 1.0, signal (assumed constant during the
> short time shown) = 0.5.
>
> The green trace shows the magnetic field as seen by a small
> tape fragment as it moves past the recording head.
>
> The red trace is the resulting magnetisation, taking the
> hysteresis effects into account. The constant value it
> converges to is the signal that remains on the tape.
>
> This is what I simulate, and it is way more complex than
> what is described in Chowdhury's paper - that would just
> be the average value of the red trace while it is still
> at full amplitude.
>
> If the signal is HF, it can't be assumed constant during
> the time the field decay takes, it may well go through a
> complete cycle or more. This is what causes the 'self
> erasure' at higher frequencies which Chowdhury doesn't
> take into account at all (AFAICS).
>
> Ciao,
>
-- 
Nicolas Bouillot [https://nicolasbouillot.net]
_______________________________________________
Linux-audio-dev mailing list -- linux-audio-dev-cunTk1MwBs/[email protected]
To unsubscribe send an email to linux-audio-dev-leave-cunTk1MwBs/[email protected]