Digest and Test: Speech Enhancement with GANs

In this chapter, we learned about the challenges in audio enhancement in speech and the advantages of the SEGAN implementation. We also learned how to implement the SEGAN model and its loss function, which combines the adversarial loss with the L1 reconstruction loss. We also learned how to train the model by writing the training routine and loading the clean and noisy data.

Then, we learned how to evaluate the model qualitatively by using TensorBoard and listening to audio samples. We also learned to evaluate the samples qualitatively by comparing spectrograms using the iZotope RX. Finally, we learned how to evaluate the model quantitatively by computing the SNR and the SSNR between the clean data and the denoised data.

Test your understanding

Put your knowledge to the test with a quiz designed to reinforce your understanding.