End-to-End Probabilistic Inference for Nonstationary Audio Analysis

Authors: William Wilkinson, Michael Andersen, Joshua D. Reiss, Dan Stowell, Arno Solin

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate, on various tasks with empirical data, how this inference scheme outperforms more standard techniques that rely on extended Kalman filtering.
Researcher Affiliation Academia 1Centre for Digital Music, Queen Mary University of London, United Kingdom 2Department of Computer Science, Aalto University, Finland 3Department of Applied Mathematics and Computer Science, Technical University of Denmark, Denmark. Correspondence to: William J. Wilkinson <william.wilkinson@aalto.fi>.
Pseudocode Yes The EP algorithm is summarised in Alg. 1.
Open Source Code Yes Matlab code for all methods: https://github.com/Aalto ML/ nonstationary-audio-gp
Open Datasets No The paper mentions using 'a dataset of 10 musical instrument sounds' for missing data imputation and generating 'synthetic data' by sampling from the generative model. For source separation, it mentions 'training the model on musical instrument notes (sources)'. However, it does not provide concrete access information (link, DOI, or formal citation with author/year) for any publicly available dataset used for training or general data access.
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, and test sets.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. It mentions memory requirements for data size ('1.2 Gb per second of data', '12.2 Mb per second of data') but not the hardware itself.
Software Dependencies No The paper mentions 'Matlab code' but does not provide specific version numbers for Matlab or any other software dependencies needed to replicate the experiments.
Experiment Setup Yes For ease of comparison, in all the real-world experiments we set D = 16, N = 3 and tune the parameters via single-sweep EP (ADF), with η = 0.75 and damping of 0.1. We use these parameters to directly compare the different inference methods (with the exception of the simulated data experiment where we use the known parameters). We use the exponential and Mat ern-5/2 kernels for κd and κg.