Text-to-speech: when AI reads text out loud
Text-to-speech turns written text into spoken audio. Thanks to AI, these voices now sound so natural that they can be hard to tell apart from a real person.
Data as of October 22, 2026
Key figures
| Level | Beginner |
|---|---|
| Origin | 2016 |
| Introduced by | DeepMind (WaveNet, voix générée par réseau de neurones) |
From text to voice in seconds
You paste an article, an email or a book chapter, and the tool reads it aloud. That’s text-to-speech, often shortened to TTS. The technology has been around for a long time, but older voices were easy to spot with their choppy, flat delivery.
The big shift came with neural networks. In September 2016, DeepMind unveiled WaveNet, a model that generates audio directly, one sample at a time, and which significantly narrowed the gap with human speech in its own listening tests. Since then, synthetic voices have gained intonation, breathing and emotion.
Where you already hear it
- Everyday life: your GPS announcing the next turn, your phone reading a web page aloud, an AI-narrated audiobook.
- With an AI assistant: voice mode lets the assistant answer out loud. It chains speech recognition (to understand you), a language model (to write the answer) and text-to-speech (to say it). The AI assistants comparison shows which ones offer voice.
- At work: narrating a training video, recording a voicemail greeting, dubbing a presentation into another language.
Smart speakers and other voice assistants work on the same principle.
The flip side: voice cloning
A highly realistic synthetic voice can imitate a specific person from a short recording. That’s the territory of audio deepfakes and some AI scams: a fake call from a relative in trouble, or a fake message from an executive asking for an urgent wire transfer.
A good habit: if you get an unusual request for money over the phone, hang up and call the person back on the number you know. Agreeing on a family code word is a simple precaution.
Common misconceptions
- « A natural-sounding voice must be a human. » Not anymore; the sound of a voice alone proves nothing.
- « The machine understands what it reads. » It pronounces text; meaning and accuracy depend on the text, not the voice.
- « A generated voice is free to use however I like. » Terms vary by service, especially for commercial use or a cloned voice.
Related terms
- Speech recognition · Technology that turns spoken words into written text: dictation, captions, voice commands and meeting notes.
- Voice assistant · An assistant you control by speaking, on a phone, smart speaker or car: it listens, works out the request and answers out loud.
- Deepfake · An AI-made or AI-altered image, video or voice recording that makes a real person appear to say or do something they never did.
- AI scams · Frauds that use AI, such as cloned voices, fake videos or very convincing messages, to steal money or personal data.
- Multimodal AI · AI that can understand or produce several kinds of content, such as text, images, audio and video, often within the same conversation.
Full glossary
Frequently asked questions
How is it different from speech recognition?
They are two directions of the same road. Text-to-speech goes from text to voice; speech recognition goes from voice to text. A voice assistant usually uses both, one after the other.
Can someone clone my voice?
Some tools can imitate a voice from a recording. Reputable services generally require the consent of the person whose voice is used. Imitating someone's voice without permission can be illegal and is also used in scams.
Does text-to-speech handle names and numbers well?
Modern voices handle everyday English well. Proper names, acronyms, numbers and foreign words remain classic stumbling blocks, so listen back before you share anything.
Is it useful for accessibility?
Yes, that's one of its original purposes. Screen readers on phones and computers rely on it to read content aloud for blind and low-vision users.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.