Impact of Minimum Hyperspherical Energy Regularization on Time-Frequency Domain Networks for Singing Voice Separation

The task of singing voice separation requires the model to maintain a trade-off between signal quality, interference introduced by music accompaniment and algorithmic artifacts. A time domain-based singing voice separation system offers a challenge in designing for low latency and in minimizing computational cost. To overcome this problem, we propose to use Gammatone auditory features for the Time-Frequency (T-F) mask-based singing voice separation task. Minimum Hyperspherical Energy (MHE) regularization in the time-domain network has recently produced the state-of-the-art result in singing voice separation (our baseline). In this work, we apply MHE to the T-F domain networks. The MHE regularized T-F domain network significantly improves the separation performance over the base-line. The MHE regularized Wasserstein Generative Adversarial Network (GAN) achieves 0.21 dB improvement in mean Signal-to-Distortion Ratio (SDR) over the baseline. Our best performing T-F domain un-regularized GAN provides an improvement of 0.75 dB and 0.63 dB in SDR over the baseline and the GANMHE, respectively. We experimentally show the failure of MHE regularized T-F domain networks with respect to their unregularized versions and have shown the need of designing a suitable adversarial objective function. We report that modifying the GAN-MHE's objective function with reconstruction loss and adapting Wasserstein GAN, results in a 0.45 dB improvement in mean SDR over its un-regularized version.