Autoregressive models
Autoregressive models use information from the previous steps and create the next output. RNNs generating text for a language modeling task is a typical example of the autoregressive model.
Figure 6.3: Autoregressive model for RNN language modeling
Autoregressive models generate the first input independently, or we give this to the network. For example, in the case of RNNs, we give the first word to the network and the network uses the first word we provided to assume what the second word would be. Then it uses the first and second word to predict the third word and so on.
Although most generation tasks are done on images, our autoregressive generation is on audio. We will build WaveNet, a research result from Google DeepMind, which is the current state-of-the-art implementation of audio generation and especially for text-to-speech processing. Through this, we'll be exploring what the PyTorch APIs for audio processing are. But before looking at WaveNet,...
