I am using AIDA-X to learn the model.
Google Colabo documentation has a 2-3 minute wav. It says,
but, input.wav distributed officially is almost 11 minutes long.
Indeed, when I used this wav file, the ESR value became much smaller and sounded great.
Could it be that the 2~3 minutes Train data is too short?
I want to eventually learn the acoustic guitar sound and make the piezo mic sound like a condenser mic.
Is it still desirable to create an 11 minute input.wav in this case?
I can’t use the official input.wav file because it is an acoustic guitar. You will have to make it yourself.
I can’t answer on the technical aspect but I also had the impression there was an increase of accuracy (and somewhat lower ESR aka Error t Signal Ratio) when using a longer file
I took the liberty of adding my own section of long chords and fast chord chugging to the file. It makes the file longer but I noticed it offered a slightly better result on some of the high gain simulations. I have yet to prove its result but since some of my models will be used in a band that plays in “drop C# tuning”, it won’t hurt adding a section of that to my file.
In dutch we say “baat het niet, dan schaadt het niet” which translate to “it might not do so much but at least it doesn’t hurt the outcome”
Speaking of acoustic guitars;
I had quite some success countering piezo “quack” by using an IR
Check this thread:
note that, on the webpage of the pedalboard I created, still has “with” and “without” reversed. the audio sample is first WITH and then WITHOUT the IR applied
.
@fxsimone the length of the dataset does effect the quality of the model you are trying to create.
The longer and diverse the better.
You could record it yourself for as long as you like. Just make sure this will also increase the training time.
The other method is to download the recordings from IDTM at https://zenodo.org/records/7544110 and build on yourself. They have acoustic guitar and one miced up in there.