Tacotron (https://arxiv.org/abs/1703.10135) was one of the first end-to-end TTS deep learning models. Basically, it is a complex encoder-decoder model that uses an attention mechanism for alignment between the text and audio. The decoder produces the spectrogram of the audio, which is then converted into the corresponding waveform by a technique called the Griffin-Lim algorithm.
The following is a representation of the entire architecture. We will present it in detail in the following paragraphs: