AI Voice for Videos: Narrator Without a Mic

An AI voice for videos comes from a text-to-speech model: you enter a script, and the model reads it out in a synthetic voice. There are three ways to get one. You pick a ready-made voice from a cloud service such as ElevenLabs or OpenAI, you clone a real voice from a short recording, or you design a new voice from a written description and generate it locally on your own computer with an open model such as Qwen3-TTS. Before you pick a tool, answer three questions: May you use the result commercially? Whose voice is it? And how will you label it as AI?

An example from my own work: for my children's book series "Mio and the Visitors", I made an audiobook with AI voices for the narrator and the characters.

Three ways to create an AI voice for videos

Qwen3-TTS calls the third way "Voice Design". Each call gives a slightly different voice, so for a series you keep one good sample as the reference and generate every further line from it.

Tools compared: prices and licenses

"Free of charge" and "free to use" are two different things with AI voices. Many open models cost nothing but forbid commercial use, some explicitly for the generated audio too. The same applies to images, see Creating AI Images: Free and Free to Use. Figures as of November 2026.

Three details lists often miss. First: if you publish audio made on the ElevenLabs Free plan, its help center says you must put "elevenlabs.io" or "11.ai" in the title, and you may not use it commercially. Second: OpenAI's documentation marks tts-1, tts-1-hd and gpt-4o-mini-tts as deprecated, with shutdown planned for January 6, 2027 and a Realtime API model as the successor. OpenAI also says its voices are optimized for English. Third: Qwen3-TTS speaks ten languages, including English, German, French and Spanish. License types are explained in Open Weight vs Open Source: Licenses.

A local model has no per-minute fee but needs computing power; the Qwen3-TTS examples are written for an Nvidia graphics card. For a Mac, see Local AI on a Mac: Mini as AI Server.

The Mio example: an audiobook with AI voices

"Mio and the Visitors" is an English-language picture book series for children. Book 1, "The Day Rumble Came to Visit", came out as a paperback on Amazon on August 17, 2026. The audiobook belongs to this book and is meant as a read-aloud video.

This is how the audiobook is built:

• Narrator and characters. A warm American narrator reads everything except direct speech. The characters only speak their own lines: Grandma Alma, the grumpy visitor Rumble, Koko the parrot, Wim the guardian and Mio himself.
• Designed, not cloned. Each voice was created from a description with Qwen3-TTS, locally on a Mac. Every further line came from the same reference, so the characters sound the same throughout.
• The text stays the book's text. Each line was generated separately and checked against the printed text with Whisper; a wrong word meant a new take.
• Listening beats measuring. The first voice for big Rumble was too scary for a children's book; now he sounds like a deep, cosy grandpa. The German narrator voice used at first spoke English with an audible accent, so there is now a separate English narrator voice.
• Sound effects instead of music. Knocking, wooden blocks and doors come from public domain recordings (CC0) and self-made sounds.
• Label in the picture. The video permanently shows "AI-generated voices", and the end card says "Voices: AI-generated".

The video was built by Eddi, my AI employee for video editing. He edits real material and makes nothing up; in an audiobook, the voice tracks are that material. He works with ffmpeg, a locally running Whisper model and Remotion. After every cut he compares two transcripts, after every encode he scans for black frames. Nothing gets uploaded without me: posting is my gate. How the book itself got onto Amazon is described in How to Publish a Book on Amazon KDP.

How to do it: five steps

1. Write a script for speaking, not for reading. Spell out numbers, abbreviations and names the way they should be spoken. Short sentences.
2. Choose the tool by license. For ads, products or a channel that earns money, every "non-commercial" model is out, and so is the ElevenLabs Free plan.
3. Pick a voice and lock it in. Generate several samples, choose one, save it as the reference.
4. Generate line by line and check twice. Whisper compares the audio word by word with your script; then you listen yourself, because no comparison measures emphasis, pace or pauses. Running Whisper locally: Transcribing Voice Recordings on Your Phone.
5. Label and approve. A short note right on the content, such as "Voice generated with AI", and the same note in the description. Publish only after you have approved it.

What the law says

This is not legal advice, just the operator's perspective, under EU and German law.

A voice belongs to its person. Under German case law, the voice is part of the general right of personality. On August 20, 2025, the Regional Court Berlin II (case 2 O 202/24) awarded a well-known German dubbing actor 4,000 euros, 2,000 euros per video, because a YouTuber had used an AI voice in two videos that listeners attributed to him. The YouTuber hadn't cloned anything: he had picked a preset voice in an AI tool, and the court treated that as if he had hired a human voice impersonator. So even a designed voice is worth a listen before you publish: does it sound like someone well known? According to the case-law database dejure.org, an appeal is pending at the Berlin Court of Appeal (case 10 U 100/25). Providers draw the same line: ElevenLabs' usage policy (last updated October 9, 2026) forbids cloning the voice of a living person without "explicit and informed consent". Recordings of anyone under 18 may not be used for a voice model there at all.

Disclosure under the EU AI Act. Article 50(4) requires deployers, which means you if you use AI professionally, to disclose that image, audio or video content is artificially generated when it is a deepfake. Under Article 3(60), a deepfake resembles real persons, objects, places or events and would falsely appear authentic. A cloned voice that sounds like a real person can fall under this. For evidently artistic or fictional works, such as an audiobook, a disclosure that does not hamper the enjoyment of the work is enough. Machine-readable marking of the output is the providers' job under Article 50(2). What applies from when is covered in AI Disclosure Rules Under the EU AI Act.

Providers require it too. OpenAI says its usage policies require "a clear disclosure to end users" that the voice is AI-generated.

Frequently asked questions

Can I create an AI voice for free?

Yes: Qwen3-TTS runs free on your own computer, and ElevenLabs has a Free plan. For commercial use, only Qwen3-TTS (Apache 2.0) is an option of these (as of November 2026).

Can I clone my own voice?

Technically yes: ElevenLabs offers it from the Starter plan, Qwen3-TTS locally from a few seconds of audio. If you use it professionally and it sounds real, it generally has to be labelled under Article 50 of the AI Act.

Does an AI voice work in languages other than English?

Yes. Qwen3-TTS and XTTS-v2 list German and other languages, and so does ElevenLabs. OpenAI says its voices are optimized for English. Test each language separately; a voice can have an accent in its second language.

Do I have to label an AI voice?

If you use it professionally, it seems real and it resembles a real person, yes, Article 50(4) of the AI Act requires it. OpenAI requires the disclosure for its voices anyway. A short label costs nothing and helps in borderline cases too.

What to do next

Take a short text you need anyway, such as the intro of your next video. First check whether you may use the result commercially, then generate three samples, pick one and compare it with your script using Whisper. Before you publish, add the note "Voice generated with AI". If you want music underneath, similar license questions apply; they are covered in the article on AI music generators. And if the voice needs a face, continue with a recurring AI character.

Eddi is available in my community, the AI Employee Framework, as a ready-made template in the Premium tier, as his own employee folder with the same checking routine he uses with me. Everything about it is on the community page.