## TL;DR
Modern TTS produces speech indistinguishable from human recordings, with voice cloning from one-minute samples and emotional expressiveness that captures laughter, whispers, and nuanced prosody.

## Core Explanation
Two-stage pipeline: text-to-spectrogram (Tacotron, FastSpeech) → spectrogram-to-waveform (WaveNet, HiFi-GAN). End-to-end models (VITS, Voicebox) unify these stages. Mel spectrograms compress audio into time-frequency representations suitable for neural processing.

## Detailed Analysis
FastSpeech 2 (Microsoft) enables parallel, non-autoregressive generation for real-time synthesis. Voicebox (Meta, 2023) approaches TTS as an in-context learning task — conditioning on a short audio sample to generate speech in any voice, including multilingual transfer.

## Further Reading
- Hugging Face: Text-to-Speech Models
- Coqui.ai TTS
- ISCA Speech Synthesis Workshop