WaveCRN: An Efficient Convolutional Recurrent Neural Network for End-to-end Speech Enhancement

This repo is an example usage of the proposed model. example/main.ipynb shows a minimal SE pipline and a visulization of the enhanced speech sample.

The architecture of the proposed WaveCRN model. For local feature extraction, a 1D CNN maps the noisy audio x into a 2D feature map F. Bi-SRU then encodes F into an restricted feature mask (RFM) M, which is element-wisely multiplied by F to generate a masked feature map F'. Finally, a transposed 1D convolution layer recovers the enhanced waveform y from F'.

torch==1.4.0
sru==2.3.5

Q:Table 1 shows that the proposed method, WaveCRN, performs slightly better than Wave-U-Net but without confidence intervals it is not possible to ascertain whether these differences are statistically significant or just happened by chance.

A: We would like to emphasize that the major advantage of WaveCRN is that it can achieve comparable performance to state-of-the-art SE methods (such as Wave-U-Net [36]) while requiring much less model complexity and computational costs. In the revised paper, we have clearly highlighted the main advantages of WaveCRN: (a) In Section III-B, we reported “It can be clearly seen from Table I that WaveCRN outperforms other models in terms of all perceptual and signal-level evaluation metrics.” (b) From Tables I and II, It is clear that WaveCRN uses a smaller size and less computational costs and achieves comparable performance to WaveCBLSTM. Table II shows that WaveCRN is 11.48/5.56 times faster than Wave-U-Net in forward/back-propagation pass in the training stage, while having only 27% of the parameter number as compared to Wave-U-Net. Due to the space limitation, we did not report the model size and the computational cost in Table II. Instead, we report Table II.R1 in our GitHub page (https://github.com/aleXiehta/WaveCRN). If you advise us to include the result, we will use Table II.R1 to replace Table II in the current manuscript.