How we made state-of-the-art speech synthesis scalable with Modular