Text-to-Speech (TTS) is the technology that converts written text into audible speech. Picture a tireless reader able to give a voice — a timbre, an intonation, a rhythm — to any sentence, in any language, within milliseconds.
How it works
Modern systems rely on deep learning and operate in two main stages:
- Linguistic front-end: the text is cleaned, normalized (turning "2026" into "twenty twenty-six") and converted into phonemes, the smallest units of sound.
- Acoustic synthesis: a neural model (such as Tacotron or VITS) produces a mel-spectrogram, then a vocoder (WaveNet, HiFi-GAN) turns it into a sound wave.
The goal is to maximize the likelihood of the audio samples $x$ given the text $t$:
$$\theta^* = \arg\max_{\theta} \; \sum_{i} \log p_{\theta}(x_i \mid t_i)$$
Approaches compared
| Approach | Naturalness | Cost | Usage |
|---|---|---|---|
| Concatenative | Medium | Low | Old GPS units |
| Parametric (HMM) | Robotic | Very low | Embedded systems |
| Neural (end-to-end) | Very natural | High | Modern voice assistants |
Uses and stakes
TTS powers voice assistants, accessibility for the visually impaired, audiobooks and dubbing. Recent models enable voice cloning from just a few seconds of recording — a powerful advance that also raises real risks of impersonation and audio deepfakes.
Giving a voice to the machine brings AI closer to the human: the challenge is keeping control over who is really speaking.