No documentation for training.
Opened this issue · 1 comments
The tutorial video repeats the same single process of playing an mp3 from the DB for eleven minutes. Training and parameters are not showcased, nor there is any detail as to which of the Essentia clustering methods (all pretty outdated) you decided to implement for your experiment.
It'd be good to have more examples to provide you with useful feedback.
Hello,
sorry for the late response. This repository is part of my bachelor thesis which I wrote two years ago. As you might have noticed the repository was not touched since then. You are right, most of it is outdated and I would definitely do things different today. Most of the detailed documentation can be found in my thesis, which is written in German, unfortunately. But still you could get an idea on which algorithms were used even if you do not speak German. Moreover you could have a look at the module documentation for feature_extractor.py and search.py for some details in English. For more details on possible usage parameters you could also try invoking the program with the -h flag.
Apart from this implementation, the main work of my thesis has been a literature overview on which features people could rely on to denote two sounds as 'similar'. This first part should help on selecting adequate features for the implemented algorithm. Moreover to evaluate the algorithm, a benchmark was build for a small set of sounds, using crowd sourcing. The algorithm was then evaluated against this benchmark and compared with and without MFCCs, to the algorithm of freesound.org and to a simple baseline using only MFCCs. As evaluation metrics accuracy, average score deviation and normalized discounted cumulative gain were used. The results where mainly that all algorithms performed quite similar with algorithms that used MFCCs generally performing better.
I would be interested in how you came across this repository and what kind of feedback you are thinking about.