Stefan Uhlich¹, Franck Giron¹, Michael Enenkl¹, Thomas Kemp¹, Naoya Takahashi², Yuki Mitsufuji²
¹Sony European Technology Center (EuTEC), Stuttgart, Germany
²Sony Corporation, Audio Technology Development Department, Tokyo, Japan
stefan.uhlich (at) eu.sony.com
- is_blind: no
- additional_training_data: yes
- Code: not available
- Demos: not available
This submission uses a bi-directional LSTM network as described in [1] with three BLSTM layers, each having 500 cells. For each instrument, a network is trained which predicts the target instrument amplitude from the mixture amplitude in the STFT domain (frame size: 4096, hop size: 1024). The raw output of each network is then combined by a multichannel Wiener filter as described in [2] where we estimate the power spectral densities and spatial covariance matrices from the DNN outputs.
The network is trained on an internal database with 804 songs. We used the
full train part of musdb as validation set, which is again used to
perform early stopping and hyperparameter selection (LSTM layer dropout
rate, regularization strength).
- [1] S. Uhlich, M. Porcu, F. Giron, M. Enenkl, T. Kemp, N. Takahashi and Y. Mitsufuji. "Improving music source separation based on deep neural networks through data augmentation and network blending", Proc. ICASSP, 2017.
- [2] A. A. Nugraha, A. Liutkus, and E. Vincent. "Multichannel music separation with deep neural networks." EUSIPCO, 2016.