Text-to-speech
In short
Text-to-speech (TTS) is technology that turns written text into spoken audio. Modern AI versions produce voices that can sound close to human, and power read-aloud apps, audiobooks and chatbot voices.
- Also called
- TTS, Speech synthesis, AI voice, Read aloud
- Topic
- How AI models work
- Level
- Beginner
A text-to-speech system reads text and outputs a voice. It is the reverse of speech recognition, which turns speech into text.
Older systems worked by rules or by stitching recordings together. A 1987 review by Dennis Klatt describes rule programs that drove a synthesiser, alongside methods that joined recorded snippets of speech, and lists the scientific problems that kept them from sounding fully human1.
Neural networks changed that. WaveNet, from Google DeepMind in London, generates raw audio one sample at a time, with each sample predicted from the ones before2. Speech needs at least 16,000 samples per second, which is why this was hard2. In tests, listeners rated WaveNet's English and Mandarin speech as clearly more natural than the best older systems2. A single WaveNet could also switch between many different speakers' voices2.
Today TTS is offered as a service by AI companies. OpenAI's API, for example, provides several TTS models, a set of built-in voices and audio formats such as MP3 and WAV3. OpenAI's usage policies require developers to tell listeners clearly that the voice they hear is AI-generated and not a human3.
TTS is not the same as voice cloning, though the two are linked. Cloning copies one particular person's voice, which raises consent and fraud risks that standard built-in voices do not.
Example
Someone with dyslexia uses a read-aloud app to listen to a long council letter or a PDF payslip instead of reading it on screen. The app is running text-to-speech.
Why it matters
It makes text accessible to people who struggle to read a screen, and it gives chatbots a voice. As voices get more lifelike, telling a recording of a person from an AI voice gets harder.
Questions people ask
Is text-to-speech the same as voice cloning?
No. Text-to-speech is any system that reads text aloud, usually in stock voices. Voice cloning is a specific use that copies one real person's voice. Many TTS services offer both, but cloning raises questions about consent that ordinary built-in voices do not.
How does AI text-to-speech sound so human?
Neural models learn directly from recordings of people speaking. DeepMind's WaveNet, for example, models the raw audio waveform sample by sample, and listeners rated it as more natural than earlier systems2.
Do I have to say a voice is AI-generated?
It depends on the service and the law where you are. OpenAI's usage policies, for example, require developers using its TTS to tell end users clearly that the voice is AI-generated and not a human voice3.
Related terms
Tools that use it
- ElevenLabs (ElevenReader and ElevenLabs app)AI voices that read books, PDFs and web pages aloud in ElevenReader, and make voiceovers or voice clones in the ElevenLabs app.
- Eleven v3Expressive ElevenLabs text-to-speech model with audio tags and multi-speaker dialogue
- SpeechifyA read-aloud app that turns PDFs, web pages, emails and photographed book pages into speech, with word-by-word highlighting.
- NaturalReaderText-to-speech reader for documents, web pages and photographed pages, with AI summaries and voice cloning.
- Gemini 3.1 Flash TTSText-to-speech model from April 2026 with inline audio tags to direct voice style