AI text to speech

Quick answer: AI text to speech (TTS) reads written text aloud in a neural voice: the model works out how each sentence should sound, its words, stress, pauses and pitch, and then produces the audio, so it sounds read by a person rather than stitched together from recorded syllables. Choose a tool by the languages and voices you need, how much control it gives over pronunciation and pace, and whether you may use the audio commercially. On SceneMixer the lines are not read over a video afterwards: each character speaks them inside the generated clip, in that character’s voice.
Still from “Neon Debt”, an AI short drama generated end-to-end by SceneMixer (7 characters, 1 episode)
SampleNeon DebtGame Engine Realism1 episodeProduced end-to-end by SceneMixer from the scriptWatch the episodes

How AI text to speech works

A modern text to speech system goes through three steps:

  1. Text normalisation: numbers, dates, abbreviations and symbols become words, and the context decides how: “Dr.” is “doctor” before a name and “drive” after a street name.
  2. Pronunciation and prosody: the words become sounds, and the model predicts stress, pauses, pitch and speed from the whole sentence, which is why a question rises at the end and a list has small breaks.
  3. Audio: a neural model produces the waveform, the sound itself, either from an intermediate picture of the sound called a spectrogram or directly.

Older systems glued together short recordings of a voice actor, which is why they sounded choppy at the joins. Neural voices are trained on many hours of recorded speech and generate every sentence fresh, so the same voice can read text it has never seen with a natural rhythm.

AI text to voice, AI voice generator, TTS: what the names mean

NameWhat it usually meansTypical use
Text to speech (TTS)Text in, spoken audio out, in a ready-made voiceNarration, accessibility, reading articles aloud
Text to voiceAnother name for the same thingThe same
AI voice generatorText to speech plus voice design: pick or describe a voice, often clone oneVoice-overs, characters, ads
Voice cloningA new voice built from a recording of a real personYour own voice at scale, with consent
AI voice changerSpeech in, speech out: your own recording re-voicedStreams, re-voicing a take

So a tool sold as an AI voice generator is usually text to speech with a bigger library and more control. The voice changer is the odd one out: it needs a recording to start from, not text.

How to choose an AI text to speech generator

  • Languages and accents: check the exact language and region you need, and listen to the voice in it; a voice that sounds natural in one language can sound foreign in another.
  • Voices: preview several with a paragraph of your own script, not the demo line, and on the device your audience will use.
  • Control: pauses, emphasis, speed, pitch, and a pronunciation list for names and brand words. Many tools accept SSML, a markup for exactly these.
  • Long text: how many characters fit in one request and in a month, and whether long text is split cleanly at the ends of sentences.
  • Rights: whether the audio may be used commercially, in ads or on monetised channels, and whether credit is required.
  • Output: the file format, such as MP3 or WAV, and an API if the voice has to run inside an app.

Free AI text to speech, or text to speech free: what free plans usually limit

Free plans are good for testing a voice, and most limit at least one of these: characters per month, the number of voices, downloads, commercial use and the quality of the file. Some free online readers only play the audio in the browser and do not let you save it. Before you build a project on a free plan, check that the voice you chose is on the plan you would pay for, and read the licence for the place you will publish.

Write text that reads well aloud

  • Spell out anything ambiguous: “2/3” could be “two thirds” or a date.
  • Use punctuation for timing: a comma is a short pause, a full stop a longer one, a new paragraph longer still.
  • Respell hard names the way they sound, or add them to the tool’s pronunciation list.
  • Keep sentences short; long ones flatten the intonation.
  • Generate one paragraph, listen, and fix the text, rather than regenerating the same words again and again.

AI voice over for video: read over it, or spoken in the scene

There are two ways to put an AI voice on a video. The usual one is a voice-over: generate the narration as audio, lay it under the picture in an editor and trim the picture to fit. It suits explainers, tutorials and documentaries, where nobody on screen speaks.

When a character on screen has to say the line, a separate audio file means syncing the lips afterwards. Newer video models generate the speech together with the picture instead, so the speaker’s mouth and the words come from the same pass. That suits drama and dialogue rather than narration.

Make your own AI voice: cloning and consent

Voice cloning builds a voice from a sample recording; more clean audio usually gives a closer match. Clone only your own voice or one you have permission to use, keep that permission in writing, and say in the description when a voice is synthetic. Many services refuse to clone well-known voices, and platforms increasingly ask creators to label realistic synthetic audio.

A five-minute test before you commit to a voice

  1. Pick the paragraph of your script that has a name, a number and a question in it.
  2. Generate it in three voices and listen on the speakers your audience uses, such as a phone.
  3. Fix how names and numbers are written, then generate again.
  4. Check the pace against the picture or the reading time you need.
  5. Read the licence for where you will publish before you generate the whole script.
Narration over the videoSpeech generated with the picture
FitsExplainers, tutorials, documentariesDrama, dialogue, characters on screen
LipsNobody speaks on screen, or synced afterwardsMove with the words in the same shot
Changing a lineNew audio, then re-edit the cutRegenerate or redub that shot
The voiceAny text to speech voiceThe voice set for that character

FAQ

Does SceneMixer have text to speech?

Not as a separate audio download. On SceneMixer each character’s lines are spoken inside the generated video, in the voice set for that character, and the speaker’s lips move with the words.

Is the voice free on SceneMixer?

The voice comes with the video rather than as a separate charge: a new SceneMixer account starts with 50 credits, video is 10 credits a second on the default tier, and matching a voice to a character costs nothing.

How do I choose a voice for a character?

On SceneMixer each character has a voice written in words, such as age, weight and pace, or a preset from the voice library that you can play before choosing.

Can I make my own AI voice?

On a paid SceneMixer plan you can upload a recording you have the right to use as a character’s voice, and it is used in every segment that character speaks in.

How do I add an AI voice over?

In a SceneMixer storyboard, start the line’s note with (voice-over): the line is heard in the shot, and nobody on screen moves their lips.

Can a line sound angry, tired or whispered?

Yes. On SceneMixer each line carries a delivery note in brackets, such as (low, steady), and it is performed that way.

Which language do the characters speak?

The project’s dialogue language, chosen in step 1 on SceneMixer. Write the lines in that language; it is what the voices speak.

Can I change a line after the video is made?

Yes. Redub on SceneMixer replaces a line in a finished clip, and its Original voice option keeps the speaker’s voice from that clip.

Can I clone a celebrity’s voice?

No. SceneMixer takes only your own voice or one you have permission to use.

Is it AI text to speech software I install?

No. SceneMixer runs in a browser, and on Android as an app; the voices are generated on its servers together with the video.

Hear your characters say their lines

Write a short scene, give each character a voice, and render one segment to hear it.

Start a project

Updated November 20, 2026