This chapter shifts focus to the generation of speech, detailing the methods used to build modern Text-to-Speech (TTS) systems. The aim is to progress from basic synthesis concepts to techniques capable of producing high-fidelity, natural-sounding, and controllable artificial voices.

You will examine the architecture and training processes for several categories of state-of-the-art acoustic models:

• Autoregressive Models: Analyzing sequence-to-sequence approaches like Tacotron and Transformer-based TTS.
• Non-Autoregressive Models: Studying parallel generation techniques for faster inference, such as FastSpeech and its variants.
• Flow-Based and GAN-Based Models: Investigating alternative generative modeling paradigms applied to acoustic feature generation.

Beyond the core model architectures, we will cover methods for:

• Modeling and controlling speech prosody (rhythm, intonation).
• Generating expressive speech with varying styles or emotions.
• Implementing voice cloning and conversion systems.

The chapter includes a hands-on practical section focused on training an advanced TTS model using a contemporary toolkit.