Skip to content
Unifying probabilistic models for time-frequency analysis. Spectral Mixture Gaussian processes in state space form.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
audio/speech
data
matlab
.gitignore
LICENSE
README.md
poster.pdf

README.md

Unifying probabilistic models for time-frequency analysis

https://arxiv.org/abs/1811.02489

Paper accepted to International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2019.

In our paper we show equivalence between probabilistic time-frequency models (e.g. the probabilistic phase vocoder) and Spectral Mixture Gaussian processes. Therefore this code serves 3 novel purposes:

  • Providing an easy way to construct more complex probabilistic time-frequency models by swapping in different kernel functions.

  • Converting Spectral Mixture GPs to state space form so we can apply Kalman smoothing for efficient inference that scales linearly in the number of time steps.

  • Hyperparameter tuning in spectral mixture GPs via a maximum likelihood approach in the frequency domain (Bayesian spectrum analysis).

matlab/ folder contains the code and example scripts.

matlab/experiments/ folder allows you to rerun the missing data synthesis experiments from the paper and produce the plots.

matlab/prob_filterbank folder contains Richard Turner's standard probabilistic time-frequency analysis code.

Reference:

@inproceedings{wilkinson2019unifying,
       title = {Unifying probabilistic models for time-frequency analysis},
      author = {Wilkinson, William J. and Andersen, Michael Riis and Reiss, Joshua D. and Stowell, Dan and Solin, Arno},
        year = {2019},
   booktitle = {International Conference on Acoustics, Speech and Signal Processing (ICASSP)}
}
You can’t perform that action at this time.