This repository has been archived by the owner. It is now read-only.
Music information retrieval using deep neural networks
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

nAud: music information retrieval using deep neural networks

For information on the research side of what was done, see the Github Pages section.

Initial setup

The files expect several things. Most important is the Keras library, which you can install via pip.

Next, make the expected directories - both in the root level of the repository. One named output, for the saved neural networks, and one named cache, which will be used to store the library. (You may wish to use a symlink to create the cache folder on an external drive with sufficient free space, if your working drive has limited storage.)

Finally, download ffmpeg and put the executable file in the root of the repository.

Genre Identification

To run the genre identification network, you'll first need to build the database. Put the local computer's iTunes library file (~/Music/iTunes/iTunes Music Library.xml or C:\Users{username}\My Music\iTunes\iTunes Music Library.xml) in the 'data' folder of the repository, and then put its filename in or (Using the 'limited' option is recommended, though you'll need to change the number of songs of each genre to use.)

Run whichever variant of buildLib you opted to use - it'll take a while to run. From there, you can input the name of the database file it generates (in the cache folder) into and run your code.

To evaluate, make sure has the name of the correct library file and network files, and then run it with the output being stored to disc. It can then be read into the spreadsheet software of your choice - the data is separated by the '/' character, though this will yield an extra column full of 'cache', as the output includes the full location of each file.

Instrument Identification

Instrument identification requires a few more directories: in the cache folder, created drum, guitar, other, vocal, and test. The first four are used for training data, the final one for evaluation data.

Delete data/database.json and, by editing, create your own database file with whatever training data you'll be using.

To train, you can run Evaluation is similar to the genre ID code: input the name of the correct library file and saved neural network in, and store the output of running the code to disc.


All of the relevant research data is stored in logs, sorted by date and time of experiment.