Updated October 2026 · Voice rights, consent and deepfake labels
Label an AI-generated voice early and plainly: a spoken line near the start, a written note in the description, and an on-screen caption for video. Use simple words that say what was generated, and keep the same label in every episode where the voice appears.
Audiences are increasingly aware that voices can be generated, and many feel misled when they find out later. A clear label builds trust and, in some places, is required by law. The practical question for a creator is how to label well: where to put the disclosure, what to say, and how to keep it consistent without making the content feel like a warning label.
This guide covers the choices in order. It assumes the label is needed, and focuses on making it clear and honest without damaging the listening or viewing experience.
Not every use of software changes the meaning of a voice. Cleaning background noise, levelling a recording or cutting a pause does not generally make the speech synthetic. A generated voice, a cloned voice, or a voice that has been changed to sound like someone else does. When in doubt, label. A short note costs little and a complaint costs much more.
Write down the rule you use, so the whole team applies it the same way. A simple rule, such as labelling any speech that was not spoken by the person shown or named, covers most cases.
A label at the very end of an episode is easy to miss, and a viewer or listener may stop before reaching it. Place a spoken or on-screen label near the beginning, before the first generated line. For a podcast, a short sentence after the intro works well. For a video, a caption that appears as the voice begins is clear.
Repeat the label if the generated voice returns in a later section, especially after a long gap. Listeners who join partway through should not miss it.
Use words that say exactly what happened. “The narration in this episode was generated by an AI voice model, based on a recording of the host” is clear. “Enhanced audio” is vague and may mislead. Avoid technical jargon, and do not use a phrase that suggests the voice is real if it is not.
Where the voice is cloned from a real person, say so, and say whether that person agreed. That detail protects both the speaker and the audience.
A written note in the episode description or the video description is a useful second layer. It gives listeners a place to check details, and it helps search and transcripts record the disclosure. The written note should not replace the spoken or on-screen label, because many people will never read the description.
Keep the wording identical across episodes, so listeners learn to recognise it and trust it.
If you publish a transcript or captions, make sure the label appears there too. Automatic captions may not mark the generated voice, so add the note by hand. A transcript that hides the label undermines the disclosure, even when the audio includes it.
Check the captions against the audio, since an error in a name can make a disclosure less clear.
A label used in an audio episode should match the label used in a video, a clip and a social post. If a clip is cut from an episode, the clip needs its own label, because the viewer may never hear the episode. Plan the label into every format from the start, rather than adding it at the end of the process.
Check whether the label still matches what you do. If you change the tool, the voice or the way it is used, update the wording and the placement. Note the date of each change, so you can show when the disclosure began and what it said.
Before you publish a series with a new label, ask someone who has not heard the episode to play it and tell you what they understood. If they miss the disclosure, or think the voice is real, the label is not clear enough. A short test catches the problem before it reaches the whole audience.
Test the label on a phone speaker as well. A line that is easy to read on a laptop can be hard to hear on a small speaker in a noisy room.
A label does not make a misleading use acceptable, and it does not replace consent. Some uses may be prohibited whatever the disclosure says, so check the rules that apply. This guide is general practice, not legal advice, and it does not describe any specific legal requirement in full.
Real voices need clean recordings just as much. The Vocal Checker shows clipping, noise and level problems in a recording, and the Tunedzy Go page covers cleaning a phone recording in your browser. For the rules that may require the label in the EU, read the EU labelling post in this series.
Spot clipping, noise, rumble and level problems. Runs in your browser, nothing is uploaded.
Near the start, before the first generated line, and repeated if the voice returns after a gap.
No. Many listeners never read it. Use a spoken or on-screen label as well.
Plain words that say what was generated and, if relevant, whose voice it is based on.
No. Noise removal and levelling do not make the speech synthetic.
