U-Net-based Speech Dereverberation with directional feature from spherical microphone array recordings
An implementation of this paper.
create.py performs following procedure:
- Calculate anechoic spherical harmonic domain (SHD) signals from speech sources and spherical Fourier transform basis
$\mathbf Y_s$ (Ys). - Calculate 32-channel spherical microphone array recordings from speech sources, room impulse responses (RIRs), and the modified inverse of rigid sphere modal strength
$b^{-1}_n(kr)$ (bEQf). - Calculate reverberant SHD signals from the result of 2.
- Perform STFT signals.
- Calculate directional features, one of spatially-averaged intensity vector (SIV) and direction vector (DV).
- Save magnitude and phase of the STFT of the 0-th order SHD signals and directional features.
Read docstring of create.py for usage.
main.py is used to train or test DNNs.
Read docstring of main.py for usage.
The DNN model is based on FusionNet (U-Net-like DNN). Refer to model directory.
Source codes for PESQ, STOI, and fwSegSNR are in matlab_lib directory.
Frequency-domain SegSNR is implemented in audio_utils.py.