Neural FCA

This package provides neural blind source separation methods, namely neural full-rank spatial covariance analysis (neural FCA) [Bando2021] and neural fast FCA (neural FastFCA) [Bando2023].

Tasks

AviTask(encoder, decoder, scm_estimator, ...)	PyTorch Lightning implementation of the AVI task of neural FCA [Bando2021].
JdAviTask(encoder, decoder, n_fft, ...)	Neural FastFCA task [Bando2023] implemented in PyTorch Lightning.

Encoders

DilcnvEncoder(n_fft, n_mic, n_src, dim_latent)	Encoder module using dilated depthwise separable convolutions for multichannel spectrograms.
UNetEncoder(n_fft, n_mic, n_src, dim_latent)	Encoder module using a 1D U-Net architecture for multichannel spectrograms.
JdUNetEncoder(n_fft, n_mic, n_src, ...[, ...])	U-Net-based encoder for multichannel audio representation learning.

Decoders

ResDecoder(n_fft, dim_latent[, io_ch, n_layers])	Decoder module that transforms latent representations into magnitude spectrogram estimates.
ResLinearDecoder(n_fft, dim_latent[, io_ch, ...])	Decoder module that reconstructs a magnitude spectrogram from a latent representation.

Lightning Callbacks

PsdVisualizerCallback()	Callback to visualize intermediate tensors during model validation.
XtVisualizerCallback()	Callback to visualize the time-frequency energy (Xt) of the model output during validation.

References

[Bando2021] (1,2)

Yoshiaki Bando, Kouhei Sekiguchi, Yoshiki Masuyama, Aditya Arie Nugraha, Mathieu Fontaine, and Kazuyoshi Yoshii, “Neural full-rank spatial covariance analysis for blind source separation,” IEEE Signal Processing Letters, vol. 28, pp. 1670-1674, 2021. [PDF]

[Bando2023] (1,2)

Yoshiaki Bando, Yoshiki Masuyama, Aditya Arie Nugraha, and Kazuyoshi Yoshii, “Neural fast full-rank spatial covariance analysis for blind source separation,” in Proc. 31st European Signal Processing Conference (EUSIPCO), pp. 51-55, 2023. [PDF]