WaveGAN

WaveGAN is a generative adversarial network designed to synthesize raw audio waveforms. It demonstrates end-to-end audio generation without relying on spectral representations or vocoders. The model comprises a generator that maps a random latent vector to a waveform and a discriminator that attempts to distinguish generated audio from real samples. WaveGAN is part of early efforts to apply GANs directly to audio signals.

The architecture uses 1D convolutional networks. The generator stacks several 1D transposed convolution layers to upsample a compact latent vector into a waveform of fixed length, corresponding to a short audio excerpt at common sampling rates. The discriminator applies a sequence of 1D convolutions to judge realism, typically with leaky ReLU activations. The training objective follows the standard adversarial loss framework.

WaveGAN is typically trained on unlabeled audio datasets, such as speech or music, and aims to produce plausible sounds rather than specific controllable outputs. The generator takes a random vector and outputs a waveform that the discriminator rates against real samples. In practice, WaveGAN can produce short clips with rudimentary naturalness, serving as a baseline for subsequent improvements in end-to-end audio synthesis.

Impact and limitations: WaveGAN demonstrated that raw waveform generation with adversarial networks is feasible, influencing later work in neural vocoders and spectrogram-based approaches. Limitations include artifacts and limited long-range coherence, leading to ongoing research into higher-capacity architectures, conditioning, and training techniques to improve realism and controllability.