In this chapter, we provided an overview of the TTS field. We explained the criteria that a good TTS system should follow. We explored the tip of the iceberg in traditional TTS methods.

Then, we presented a state-of-the-art, end-to-end deep learning approach, Tacotron. Its implementation concluded the chapter, with instructions to experiment with the model on an open source dataset adapted to the problem.   

In the third part of the chapter, we saw how Keras simplifies the process of building seemingly complex neural networks. However, making a prototype is one thing; scaling up is another thing. Indeed, even if a proof of concept runs smoothly on your computer, managing to make it functional for a large amount of users on different platforms (web and mobile) can be challenging, for many reasons (such as throughput). The next chapter will tackle the problem of shipping deep learning models to production environments.