Naoya Takahashi¹, Nabarun Goswami², Yuki Mitsufuji¹
¹Sony Corporation, Audio Technology Development Department, Tokyo, Japan
²Sony India Software Center, Bangalore, India
Naoya.Takahashi [at] sony.com
- is_blind: no
- additional_training_data: yes
- Code: not available
- Demos: not available
This submission uses a multi-scale multi-band DenseLSTM (MMDeseLSTM), which is an extetnsion of MMDenseNet [1]. In MMDenseLSTM, some of dense-blocks, typically in low scale, equip a LSTM layer, as shown in Figure 1. Each LSTM layer is one layer bi-directional LSTM with maximum 128 cells.
For each instrument, a MMDeseLSTM is
trained to predict 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.
More details, experiments and analysis is described in a paper which is submitted to elsewhere.
We used the train part of musdb and an internal database with 804 songs as development set and randomly chose 48 songs as validation set, which is used to
perform early stopping and hyperparameter selection (growth rate,
number of layers).
- N. Takahashi and Y. Mitsufuji: Multi-scale multi-band DenseNets for audio source separation, Proc. WASPAA, 2017
- A. A. Nugraha, A. Liutkus, and E. Vincent. "Multichannel music separation with deep neural networks." EUSIPCO, 2016.
- N. Takahashi, N. Goswami and Y. Mitsufuji: MMDenseLSTM: an efficient two way modeling for audio source separation, arXiv pre-print, 2018