First off, what is a GAN? It stands for Generative Adversarial Network, which basically means we have two neural networks fighting each other to create something new and awesome. In this case, one network (the generator) tries to generate fake audio waveforms that look like real ones, while the other network (the discriminator) tries to figure out whether a given audio sample is real or fake. The goal of training is for the generator to fool the discriminator into thinking its generated samples are actually real.

Now how we train this thing using WaveGAN. First, you need some data in our case, we use MP3 files (or WAV files if you prefer) as input. We then preprocess these audio clips by extracting 16kHz raw waveform samples and converting them to a format that can be fed into the neural network. This is where things get interesting: instead of using traditional image-based GANs, we use WaveGAN to generate raw audio waveforms directly from scratch.

To train our model, we first load up some data and split it into training and validation sets (you can do this manually or let the script handle it for you). We then initialize our neural network with some random weights and start feeding in batches of audio samples to both the generator and discriminator networks. The generator tries to generate fake waveforms that look like real ones, while the discriminator tries to figure out whether a given sample is real or fake.

During training, we use a technique called backpropagation to update the weights of our neural network based on how well it performs at generating and distinguishing between real and fake audio samples. This involves calculating gradients (which are like little arrows pointing in the direction that will help us improve performance) for each weight in the network, and then using these gradients to adjust those weights accordingly.

After a certain number of training iterations (usually around 10-20 epochs), we save our model’s current state as a checkpoint file so we can resume training from where we left off if needed. We also log various metrics such as the loss function and accuracy to help us monitor progress over time.

In terms of companies and researchers in this space, there are several prominent players working on AI sound generation using GANs. One notable example is WaveGAN, which was created by a team at Google Research in 2019. This model can generate one-second-long sound effects and speech snippets that are very difficult to distinguish from real audio samples. Other researchers have also explored the use of GANs for music generation and synthesis, with some promising results in this area as well.