HAMR 2013 Proceedings

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== Source Separation for Beat Tracking ======

| Authors | Hunter McCurry |
| Affiliation | The Echo Nest |

My weekend goal was to explore the use of HPSS (Harmonic / Percussive Source Separation) to determine whether it would be a valuable pre-processing step for attribute analysis. For the purposes of time, I limited myself to looking at the differences in quality of the Echo Nest's beat-tracking algorithm between the analysis of unprocessed audio and analysis of just the percussive component of the audio.

===== - HPSS =====

I was looking for a simple and fast Harmonic / Percussive source separation method that could be implemented as my an analysis pre-processing step. The idea is to take the spectrogram and filter out high-frequency variation in each the X (time) and Y (frequency) dimension to achieve the harmonic and percussive components, respectively.

==== - Median Filtering ====

Derry FitzGerald has a paper called Harmonic/Percussive Source Separation Using Median Filtering that uses Median Filtering (a relatively straight-forward image processing technique) to do this.((http://dafx10.iem.at/proceedings/papers/DerryFitzGerald_DAFx10_P15.pdf))

==== - LibROSA Implementation ====

Brian McFee wrote an implementation of this algorithm and included it in his LibROSA library.((https://github.com/bmcfee/librosa))

==== - [iPython demo] ====

==== - [Audio examples] ====

===== - Beat Tracking =====

I used a dataset of 102 audio files (30-second samples) for which I had human-tagged truth values in the form of beat locations. I ran the Echo Nest analyzer (version 3.2.0) on the unaltered audio files and subsequently on the percussive component of my HPSS.

==== - Evaluation code ====

Turns out it is not easy to evaluate beat-tracking algorithms with a single, simple metric. There are many proposed algorithms for doing this. In the paper Evaluation Methods for Musical Audio Beat Tracking Algorithms((Matthew E. P. Davies, Norberto Degara Aand Mark D. Plumbley)) is a comparison of 10 different methods of comparison.

===== - Results =====
HAMR 2013 Proceedings

User Tools

Site Tools

Page Tools
