The short answer
Text-to-speech (TTS) is AI that turns written text into spoken audio in a natural-sounding voice. Modern TTS models read tone and context, so they can whisper, pause or sound excited when the text calls for it. Leading examples in November 2026 are ElevenLabs' Eleven v4, launched on 28 September 2026 with 90+ languages, and Google's Gemini 3.8 Flash TTS, which supports more than 130 languages and two speakers per request.
| Plain meaning | Software that reads text aloud in a human-sounding voice |
|---|---|
| Leading models | ElevenLabs Eleven v4 and v4 Turbo; Google Gemini 3.8 Flash TTS and Flash-Lite TTS |
| Languages | Eleven v4: 90+. Gemini 3.8 Flash TTS: 130+ |
| Style control | ElevenLabs audio tags such as [whispers]; Gemini style instructions and inline tags such as |
| Main catch | Cloning a voice needs consent, and the EU AI Act's transparency rules cover synthetic audio |
What text-to-speech is, in plain English
Text-to-speech, usually shortened to TTS, takes words on a page and produces audio of a voice saying them. Older systems stitched together recorded sounds and sounded flat. Today's TTS models are generative AI: they have learned from large amounts of speech how people pace sentences, stress words and change tone, and they generate new audio for any text you give them.
The opposite job, turning speech into text, is speech-to-text or transcription. The two are often sold together. ElevenLabs, for example, offers Eleven v4 for speech and Scribe v2 for transcription, and its new Eleven Realtime models, launched in October 2026, go straight from speech to speech for voice agents.
How AI text-to-speech works today
You give the model text, choose a voice and, increasingly, tell it how to perform. The two models most people will meet show how much control you now have.
ElevenLabs Eleven v4
Eleven v4 launched on 28 September 2026 and covers 90+ languages, with up to 10,000 characters per request. It follows bracketed audio tags written into the text, such as [whispers], [sighs], [laughs], [excited] or [short pause], and the Enhance button in the ElevenLabs app suggests tags for you. For dialogue, you write lines starting "Speaker 1:" and "Speaker 2:". Eleven v4 Turbo is the real-time version, with median inference latency of around 100 milliseconds according to ElevenLabs, which is what voice assistants and phone agents need.
Google Gemini 3.8 Flash TTS
Google's speech generation documentation, updated on 28 October 2026, lists Gemini 3.8 Flash TTS (more than 130 languages) and Gemini 3.8 Flash-Lite TTS (more than 100). There are 30 prebuilt studio voices plus a larger voice library. You set delivery with a style instruction such as "whispered urgently" and add momentary sounds with inline tags such as <laugh>. One request supports up to two speakers with prebuilt voices, which suits a two-person explainer.
Voice cloning
Cloning makes a TTS voice that sounds like a specific person. ElevenLabs' Instant Voice Cloning works from about one to two minutes of clean audio and asks you to confirm you have the right and consent to clone the voice; it starts on the Starter plan. Professional Voice Cloning needs at least 30 minutes of audio, ideally two to three hours, requires voice verification and only works for your own voice. Our guide to cloning your voice with AI covers the steps.
Why text-to-speech matters at work
- Voiceovers without a studio. Training videos, product demos and explainer clips can be voiced in an afternoon, and re-voiced when the script changes. See the best AI for voiceovers.
- One script, many languages. With 90 to 130+ languages available, a team can publish the same onboarding module in every market. Have a native speaker check names, numbers and tone before release.
- Accessibility. Audio versions of reports and policies help colleagues who prefer listening or who have visual impairments.
- Voice agents. Low-latency models power phone and support agents. If you are planning one, read the best AI for customer support.
- Avatar video. Tools such as Synthesia and HeyGen pair TTS with an on-screen presenter, and Descript lets you fix a recorded line by editing the transcript.
The rules you need to know
The European Commission's AI Act timeline says the Article 50 transparency rules apply from 2 August 2026. They include disclosure duties such as telling people they are talking to a chatbot and labelling deepfakes, and providers of systems that generate synthetic content and were already on the market before 2 August 2026 have a transitional deadline of 2 December 2026. In practice: tell listeners when a voice is synthetic, never clone a colleague or customer without written consent, and keep that consent on file. Our EU AI Act explainer has the full timeline.
Privacy matters too. ElevenLabs uses some of the data you provide to improve its models unless you opt out: click your profile icon, then Terms and privacy, then Data use, switch off "Improve the models for everyone" and click "Update your choice". Enterprise data is excluded by default. Do this before you upload a confidential script.
An example you can try in ten minutes
ElevenLabs has a free plan with 10,000 credits a month, and text-to-speech costs about one credit per character, so a short test costs nothing. The free plan has no commercial licence and no voice cloning, so use it for testing only.
- Take a real paragraph from your work, such as the first 80 words of a product update.
- Generate it once as plain text with a stock voice.
- Now add two or three audio tags where a person would change tone, for example [excited] before the good news and [short pause] before the call to action, and generate it again. Or press Enhance and let the app suggest tags.
- Listen to both with a colleague, without saying which is which, and ask which sounds more human.
- Check every number, name and acronym. TTS models still mispronounce product names; spelling them phonetically in the script often helps.
For a full walkthrough, see our ElevenLabs guide and the best AI for podcasts.
Related terms worth knowing
- Multimodal AI: models that take in and produce audio, images and video as well as text.
- Generative AI: the family TTS belongs to.
- Inference: running the model; latency is why real-time voice needs special "Turbo" or "Flash" models.
- AI agent: voice agents combine TTS, transcription and a language model.
- EU AI Act: sets the transparency duties for synthetic voices.
Common misconceptions about text-to-speech
- "AI voices still sound robotic." Current models handle emotion, pauses and multi-speaker dialogue. The giveaway is now usually mispronounced names or odd emphasis, not a robotic tone.
- "If I can upload it, I can clone it." ElevenLabs makes you confirm rights and consent for Instant cloning and restricts Professional cloning to your own verified voice. Cloning someone without permission can break platform rules and the law.
- "The free plan is fine for client work." ElevenLabs' free plan has no commercial licence. Check licence terms on any tool before publishing.
- "Nobody can tell it is AI." ElevenLabs runs a public AI Speech Classifier to check whether audio came from its models, and EU law expects synthetic content to be disclosed. Plan to label it.
- "One voice works in every language." Coverage is wide, but quality varies by language and voice. Test in each market's language with a native speaker.
Prompts to copy
Rewrite a script for the ear
Rewrite this text so it sounds natural read aloud by an AI voice: [paste text]. Use short sentences, spell out numbers and acronyms the way they should be spoken, write product names phonetically in brackets the first time, and mark where to pause with [short pause]. Keep it under [90] seconds at a normal speaking pace.
You get: A script a TTS model reads cleanly first time
Draft a voice consent note
Draft a short, plain-English consent form for a colleague who agrees to let our team create an AI clone of their voice for [purpose]. Cover what the voice will be used for, where it is stored, who can use it, how long we keep it and how they can withdraw consent. Flag anything a lawyer should review.
You get: A starting point for written consent before any cloning