description.md

RGT2

Gerard Roma, Owen Green, Pierre Alexandre Tremblay University of Huddersfield [email protected]

Additional Info

• is_blind: no
• additional_training_data: no

Supplementary Material

• Code: https://github.com/flucoma/LVA-ICA-2018/
• Demos: http://www.flucoma.org/LVA-ICA-2018/

Method

This system employs a Convolutional Neural Network with fully-connected output layers. The input of the network is a slice of 11 STFT frames (about 200ms). The output is a binary mask corresponding to one spectral frame. We trained the network by optimizing the negative log likelihood loss from a 2D softmax output layer. The target vector was encoded with class labels corresponding to the source with highest magnitude for each time-frequency bin.

References

• G. Roma, O. Green, P.A. Tremblay, Improving single-network single-channel separation of musical audio with convolutional layers. Proceedings of LVA/ICA, 2018