The code for multi-channel speech enhancement (and source separation) which will be presented at EUSIPCO2019. The latest version is available in https://github.com/sekiguchi92/SpeechEnhancement.
- FastFCA is a method for general source separation. In fact, it can be available only for speech enhancement because of the strong initial value dependency.
- FastMNMF is a general source separation method which integrate NMF-based source model into FastFCA.
- FastMNMF-DP is a method which integrates deep speech prior into FastMNMF, and is for speech enhancement.
- Tested on Python3.6
- numpy
- pickle
- librosa
- soundfile
- progressbar2
- chainer (6.1.0 was tested) (for MNMF-DP, FastMNMF-DP, ILRMA-DP)
- cupy (6.1.0 was tested) (for GPU accelaration)
python3 FastMNMF.py [input_filename] --gpu [gpu_id]
Input is the multichannel observed signals.
If gpu_id < 0, CPU is used, and cupy is not necessary.
If you use my code in a research project, please cite the following paper:
Kouhei Sekiguchi, Aditya Arie Nugraha, Yoshiaki Bando, Kazuyoshi Yoshii:
Fast Multichannel Source Separation Based on Jointly Diagonalizable Spatial Covariance Matrices,
arXiv preprint arXiv:1903.03237, 2019