AI voice-over for videos: text to speech that sounds right

When is an AI voice-over the right choice?

When the words matter more than who says them. A synthetic voice is consistent, quick to change and costs no studio time, so it suits some jobs well and others badly:

The rule of thumb: if the audience should feel something because of how a line is said, use a person.

How do you add an AI voice-over to a video?

1. Write the script for the ear. Short sentences, one idea each, with punctuation where you want a breath.
2. Generate each line separately, so each is easy to replace.
3. Pace it to the picture. Each line should end before the shot it describes does. Change the speed or the wording, not the picture.
4. Give it the room. A dry voice over a scene recorded in a room sounds pasted on. Match its tone and add the room's echo.
5. Duck the music under it and bring the music back up between lines.
6. Declare it. Say in the description that the voice is synthetic.

In Sharpy, the agent's speak tool reads a line in one of 54 voices, at a speed from half to twice the normal pace, and returns the audio with each word's timing, so the captions land on the words. On its maker's test, eight English lines came back word for word when Sharpy transcribed its own speech: none of 65 words wrong. It takes about a fifth of a second of work for every second of speech on the base Mac mini M4 with 16 GB, so a whole narration is quicker to make than to listen to.

How do you write a script for text to speech?

Write it the way you want it said, not the way it looks on a page:

• Spell out numbers, symbols and abbreviations as words: "twenty twenty-six", "per cent", "A I". Sharpy refuses anything with no Latin letter or digit, such as an emoji or a word in another script, and names it, so nothing is silently mangled.
• Watch names and new words. A voice guesses at a name it has never seen. Sharpy says which words it only guessed, so you can respell them ("Sharpy" as "Shar-pee", for instance) and hear them again.
• Use punctuation as direction. A full stop is a pause, a comma a shorter one. Split long sentences: a synthetic voice never runs out of breath, so the listener does.
• Read it aloud yourself first. Anything you stumble over, the voice makes flat.

How do you pace a voice-over to the picture?

Give each line the time its shot allows and no more. If a line runs long, cut words first, then speed it up a little; a voice slowed to fill a gap sounds wrong.

Sharpy does this line by line. Its fit_line tool takes a new reading of a line, whether a re-recording, the speaker or Sharpy's own voice, and fits it to the original: each word starts where the original's did, the pauses absorb the difference, and the take is matched to the original's loudness, tone and room before it replaces the line on the track in one step. It only ever shortens speech where the original allowed less time, never slows it, and it loses no words unless you say how many it may. On one sentence its maker said in 3.9 seconds, Sharpy's voice read at three paces, from 5.4 to 6.5 seconds long, and each fitted into the 3.9 seconds and still read back word for word.

It also measures how long the original room rings after each word and gives a dry take the same room, so the new line does not sound pasted in. Under the voice, music rides 6 dB down on every line and back up between; mixing sound in Sharpy covers the rest of the mix.

Can AI voice an audio description?

Yes, and it is one of the best uses for it. Audio description tells blind and partially sighted viewers what the picture shows, in the gaps between the words, as subtitles and SDH do for viewers who cannot hear. Sharpy writes a draft of it, labelled as a machine's, in the present tense and only about what is seen, at no more than 160 words a minute; read it before you use it.

Then it speaks each description in its own voice and dips the programme under it by the dip levels in the EBU's guide to audio description (TR 084). It writes a described movie with the original picture and sound untouched, a second audio track tagged as audio description so players list it, and the voice alone as a stem. A description that cannot fit its gap is refused with the seconds it needs, so you shorten it rather than speak over the dialogue. On a 3 minute 41 second reel, three descriptions came back word for word, at 156, 160 and 159 words a minute.

Is using an AI voice illegal?

No: a synthetic voice is legal in itself. What can break the law is using one to make a real person seem to say something they did not, or to deceive.

• Other people's voices. In the United States, many states protect a person's name, likeness and voice, and the US Copyright Office's July 2024 report on digital replicas recommended a federal law against unauthorised ones. Never clone a voice without its owner's permission.
• Disclosure in the EU. The EU AI Act's transparency rules (Article 50) apply from 2 August 2026: AI-generated audio has to be marked as such in a machine-readable way, and a deep fake has to be disclosed.
• Platforms. YouTube asks creators to disclose realistic altered or synthetic content, such as making a real person appear to say something they did not; its examples that need no disclosure include cloning your own voice for voice-overs or dubs.

Sharpy builds this in. Its voice is always declared: the tool names it, and a piece that carries it says so in its description and in the file's Content Credentials. Overdub, where you retype a word and Sharpy says it in the speaker's own voice with the mouth matched, works only with the speaker's recorded permission, and the export carries Content Credentials and an invisible watermark marking it as altered.

Sources

Checked on 18 October 2026.

• Apple: software licence agreement for macOS Tahoe 26 (section F, voices)
• EU AI Act, Regulation (EU) 2024/1689, on EUR-Lex
• EU AI Act Explorer: Article 50, transparency obligations
• YouTube Help: disclosing use of altered or synthetic content
• US Copyright Office: Copyright and Artificial Intelligence, Part 1: Digital Replicas (July 2024)
• EBU TR 084: audio description dip levels