Skip to content

dhunstack/music-information-retrieval

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Exploration of MIR Problems

In this repository, I have compiled some common MIR problems that I've worked on. The list below mentions all the tasks dealt with in this repository.

Instrument Recognition

The dataset contains sounds from 8 different sources - clarinet, electric guitar, female singer, flute, piano, tenor saxophone, trumpet and violin.

I compute time domain, frequency domain and cepstral domain features for the given audio files. Used KNN for the classification problem to obtain an F1 score of 0.494.

Finally used XGBoost with sample weighting adjusted for unbalanced classes, to obtain an F1 score of 0.668.

Beat Tracking

We work with the GTZAN-genre mini dataset containing beat annotations for 100 tracks.

We predict the beats using classical signal processing approach, and compare it against the RNN implementation of madmom library.

We further calculate F1-score, cemgil and continuity metrics to evaluate beat detections and analyse the common errors.

Pitch Tracking

We work with the MedleyDB-Pitch dataset containing pitch annotations for 103 tracks.

We predict the pitches using crepe model, that works directly with the waveforms.

We automate the process of determining voicing threshold for crepe, striking the right balance between voicing recall and pitch precision.

About

Explore common MIR problems

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published