Frequently Asked Questions

Can penthy evaluate non-music audio, like podcasts, audiobooks etc.?
penthy was designed, trained and tested for music. Her performance for other audio is uncharted. Feel free to explore the possibilities.

What does the name mean?
penthy is short for Penthesilea, a skilled queen of the Amazons who fought in the Trojan War, according to Greek mythology.

What file formats are supported?
Flac and wav of any sample frequency and bit depth.

Is the result always, certainly, absolutely right?
Absolutely not!

How accurate is penthy?
Approximately 90%.
False negatives (genuine files classified as transcoded) are more common than false positives (transcoded files classified as truly lossless), especially for songs that lack higher frequencies.

Is it really free?
Yes! Free as in freedom. Check the source code here.

If a file is evaluated as truly lossless, is it identical to what came out of the studio?
Not necessarily. penthy inspects only the possibility of mp3 transcoding! Even if a file is verified as truly lossless in terms of mp3 transcoding, it could still be transcoded from a different format (like Vorbis ogg), upsampled or altered in other ways.

Will you keep my file and my personal data?
The website deletes your file automatically, after the processing ends. There are no cookies, no analytics and no personal data harvesting.

Who made penthy?
Achilleas Papastamatiou (gioypi) developed penthy as his thesis at the University of Thessaly, Greece. Professor Vaggelis Spyrou was the supervisor and provided technical assistance.

How does it work?
A convolutional neural network was trained with the highest frequencies of several songs from different genres in the form of spectrogram images. The songs were split into segments of 8 seconds and given to the network as inputs. The dataset contained the truly lossless versions of the songs and their fake counterparts – flac files transcoded from mp3 files generated from the originals. The network was created in Python 3 with Keras and TensorFlow. FFmpeg was used to extract the spectrograms.

Does the neural network learn from uploaded files?
No. The website uses an already trained model that does not evolve.

Are there any easter eggs hidden in this website?
Do you like tea? I don’t, but I don’t brew coffee either.

Can I use the result evaluation of penthy commercially?
As long as you respect the applicable laws and you are clear of copyright issues with the content you upload, yes. Still, there is no guarantee that the evaluation is correct.