Singing Voice Separation: Contributions and further work

During the project, two main contributions have been added to the official repository of DeepConvSep implementation (https://github.com/MTG/DeepConvSep): The output of the implementation contains four .wav files with each source estimate. We ran the model over the Test part of DSD100 (half of the dataset), since the model was trained using the other half (Dev files). The … Continue reading Singing Voice Separation: Contributions and further work

Singing Voice Separation: Data augmentation and robustness

Data augmentation and robustness tests were out of the scope of this project since they involve retraining the model, which is computationally demanding. Here we collect some tips for further work in this direction: Time shifting the sources: a new dataset for training could be created by circular shifting all sources or some of them … Continue reading Singing Voice Separation: Data augmentation and robustness

Singing Voice Separation: Results

Table 1 shows the results of our work. Metric values show separation performance for singing voice only, since other sources are over the scope of this report: We can compare these results with published results of the same task in [1] and in 2016 SiSEC MUS task collected in www.sisec17.audiolabs-erlangen.de Perceptual results from listening to separated tracks are … Continue reading Singing Voice Separation: Results

Singing Voice Separation: Evaluation Approach

In the singing voice separation task, the output is a set of estimated voiced and background sources. This is compared with the actual singing voice source to determine accuracy of the algorithms. MIREX defines this using the following decomposition for each sound source s_estimated(t) = s_target(t) + e_interf(t) + e_noise(t) + e_artif(t) s_target is an allowed … Continue reading Singing Voice Separation: Evaluation Approach

Singing Voice Separation: Selected Approaches

After our Literature Survey, we have selected 2 approaches to investigate that cover the different approaches to Source Separation. Here is a summary of the approaches and their online implementation sources. 1. Flexible Audio Source Separation ToolBox: This toolbox was created to provide a general audio source separation framework based on a library of structured source models that … Continue reading Singing Voice Separation: Selected Approaches

Singing Voice Separation – Literature Review

Hi! This week we searched for previous literature on SVS and the current State of the Art. The Singing Voice Separation task in the context of MiREX has appeared only since 2014. Prior to 2014 singing-voice separation systems could be classified into two categories: 1) Supervised systems : a) Maps signals into a feature space … Continue reading Singing Voice Separation – Literature Review