The network architecture related to dual-path RNN TasNet.

[tky823]

I tried reimplementation of dual-path RNN TasNet reading your paper.

DUAL-PATH RNN: EFFICIENT LONG SEQUENCE MODELING FOR TIME-DOMAIN SINGLE-CHANNEL SPEECH SEPARATION

Is the general structure published in this repository the same as dual-path RNN TasNet? I understand this repository is written for TAC. There seem to be some improvements like gated outputs. Are these modules included in dual-path RNN TasNet? https://github.com/yluo42/TAC/blob/96640a803b8193a7a507652c4a5693e57da03cbd/FaSNet.py#L15-L21

[yluo42]

No it's not completely the same as the problem definition of TasNet and FaSNet are different: TasNet attempts to estimate multiplicative masks on the encoder outputs, while FaSNet attempts to estimate the time-domain beamforming filter coefficients. The gated output layer is only applied for FaSNet to ensure the dynamic range of the estimated beamforming filters is between -1 and 1, but empirically removing the gating layer or even using a linear output layer should have a similar (or same) performance.

In DPRNN-TasNet the output layer is simply self.output without the gating layer, and the nonlinearity for the output layer can be either Sigmoid or ReLU (empirically ReLU might be better).

[tky823]

Thanks! I will read the FaSNet paper. So, the network architecture of DPRNN-TasNet is like

class DPRNNTasNet(...):
    def __init__(...):
        ...
        self.encoder = nn.Conv1d(...)
        self.enc_LN = nn.GroupNorm(...) # Layer normalization
        self.BN = nn.Conv1d(...) # Bottleneck convolution.
        self.DPRNN = DPRNN(...) # includes nn.PReLU() and nn.Conv2d(...) for mask estimation
        self.output = nn.Sequential(nn.Conv1d(...), nn.ReLU()) # empirically better
        # or self.output = nn.Sequential(nn.Conv1d(...), nn.Sigmoid())
        self.decoder = nn.ConvTranspose1d(...)

Do I understand your explanation correctly?

[yluo42]

Yes it is something like that.