This repository contains work from a project done for the EL2320 course (Applied Estimation) in KTH. The report for the project is under the report directory.
- Python 3.9
In this approach an XGBoost classifier is trained to detect beats from the audio using spectral features and a particle filter uses the XGBoost output as observations and tracks the beat location and the tempo period in a song.
We perform STFT with a window size of 0.1s and a hop size of 0.0375s, which means that we detect beats within audio blocks slightly larger than than 16th note sized blocks, with a resolution corresponding to 32th notes, for a maximum tempo value of 200 BPM. A hann window is used with STFT. We then split the frequency bands to the ranges (50,120), (120,300) and (300, 22050) and calculate the sum over those frequency bands as features. XGboost then takes as input those sum values for 32 time steps, which corresponds to 96 values in total. The prediction of XGBoost infers whether there is a beat or not at the center of the last STFT frame.
We extract the note onsets from the available MIDI files and perform tempo tracking using the method proposed in this paper, using a particle filter to estimate the switching state variable.
We also evaluate the method without a particle filter estimation of the switching state variable, but by using the optimal switching state value that minimizes the residual error between the observed and predicted onset at each timestep.
Current state: The particle filter method does not work properly or is simply not performing well on the dataset we evaluated it on. Using the single switch variable value minimizing the observed/predicted onset residual gives an Accuracy2 performance value of 83.6% on the GMD dataset described below.