Learning and dataset generating

[yakuv0]

Did a run on a true flac file (rencoded with ffmpeg) and returned mp3 as a result, this was using the model available in the release of this repo, would it be possible to include more information on testing and learning process in documentation or this ticket?!
And what does noise stand for in the learning, testing and maybe the authentication process as well?!

regards

[Sg4Dylan]

Did a run on a true flac file (rencoded with ffmpeg) and returned mp3 as a result, this was using the model available in the release of this repo

Sure.

would it be possible to include more information on testing and learning process in documentation or this ticket

Please tell me exactly what you want to know

And what does noise stand for in the learning, testing and maybe the authentication process as well?

The noise generation method does not come from any standard, but from another my project for fake lossless audio generation.
FLAD/blob/main/utils.py#L73
EmiyaEngine/blob/master/core/akkomode.py#L31

You can use FLAD to detect AAC/MP3/Opus audio separately, and find images of the feature patterns of these lossy encoding formats in the temp directory, then simply identify them with the naked eye.
So, this model just replaces your eye.
The feature pattern of the compressed format can be significantly blurred by adding random noise, so it is necessary to add noise samples during training to improve the generalization performance.

[yakuv0]

@Sg4Dylan apologies but you reply for the first part was sure for a case that might be considered as a bug where FLAD returns false value for true flac file (tst2_true.flac in tst.zip), is that how the model was trained ?!

performance section of documentation mentions 200 samples per category for test set, what are the categories and should i include the files in its original format/codec (flac,acc and mp3) or have them all rencoded to flac? is it a necessity to train the model with the noise generation method?

i understand that the noise generation method is producing frequencies to pass the 20khz and have the file appear to be true flac file, specially that FLAD doesn't rely on frequency cut off for the authentication process, this method is close to cases where its an upsampled lossy file with a variable/dynamic bitrate.
spectrogram from file (tst1_upsampled.flac in tst.zip):

log in tst.zip summary:

Final result: mp3 , File: tst0_upsampled.flac
Final result: mp3 , File: tst1_upsampled.flac
Final result: mp3 , File: tst2_true.flac

tst0_upsampled.flac: lossy upsampled constant bitrate
tst1_upsampled.flac: lossy upsampled variable/dynamic bitrate
tst2_true.flac: lossless

note: this set of files is for the eval method only, without training or test and using the release model.

[Sg4Dylan]

About the training and testing process:
All audio samples are taken from a handful of confirmed lossless audio files, resampled to 48kHz 24bit FLAC by ffmpeg using soxr with high precision, and then converted to lossy audio format.
The test produced 200 samples for each format (including true FLAC) through the same process.

have them all rencoded to flac?

In fact, lossy audio is not re-encoded to FLAC.
Because FLAC is a lossless audio format, even if it is re-encoded, it does not change its spectral features.

** About the samples/logs in tst.zip: **
You may have encountered a bug in FLAD where not all temporary files from the previous test were correctly covered due to the small length of your sample.
Also, because your sample length is very small, it needs to be judged in conjunction with visual observation.