Project
In this project, I study the music genre classification.
The main code file is in FinalProject.ipynb.
I used keras(tensor flow) to train the data from GTZAN which provides a great data site for music analysis.
GTZAN data set could be downloaded from here:
http://marsyasweb.appspot.com/download/data_sets/
The original data set was using “.au” audio file, I used footbar2000 to transform them to “.wav”
The packages that required for this project were used for this course.
I found some sound processing python code that would help me to process the data that I download from GTZAN. audiofile_read.py rp_extract.py wavio.py
The result is ok compare to what have existed online.
Something else may work: The human reaction for a music genre is mostly base on the instruments. So to improve the accuraccy, one way is to extract instrument frame out of spectrum. For example, I can replace the low requency mel frame to drum play, so '1' for 1 beat, 0 for else. Ignoring the contiune vibration sound, this might clear out the interferes and increase the accuracy.
Cite: https://github.com/tuwien-musicir/DL_MIR_Tutorial by Thomas Lidy