Automatic Instrument Identification in Polyphonic Music

Course project for DS-GA1003 Machine Learning.

Members:

• Jiyuan Qian
• Tian Wang
• Peter Li

Dataset:

MedleyDB, which is available from http://medleydb.weebly.com/

Results

We compared Convnets trained on raw audio, MFCC and CQT with traditional MIR methods that extracts Gaussian features from MFCC and its first and second order deltas. Convnets trained on handcrafted features can outperform traditional methods, and that trained on raw audio, though takes much longer training time, can achieve arguably better performance.