His son copied his late father's voice with AI to make a podcast.

Sir Michael Parkinson, a British host known for his interviews, wanted to launch a podcast. Born in Barnsley, he hosted some of the great figures of the XXᵉ century over a career spanning seven decades. He died in 2023, at 88, before he could bring this project to life himself.

His son, Mike Parkinson, subsequently agreed to let an AI company use his father's voice, with the family's consent. The BBC devoted an article to the project, in which Michael Parkinson's son defends new AI podcast discusses this decision.

The podcast is called Virtually Parkinson. It has eight episodes and features a synthetic version of Sir Michael's voice conducting new interviews. Mike Parkinson says he requested clear information for listeners, describing it as an "AI iteration" rather than an actual recording.

The BBC tests listeners' ears

The BBC tested listeners' ability to distinguish Sir Michael's AI voice from a human voice. Two audio clips were presented. One came from a human recording. The other used the AI-generated voice in Virtually Parkinson.

The result is unsettling. A majority of listeners failed to identify which of the two clips featured the human voice. Voice cloning is rapidly becoming more realistic. In some cases, listening alone provides too few clues.

Mike Parkinson himself said he was "astonished" by the accuracy of the reproduction of his father's voice. He also emphasized the framework chosen for this podcast. The family gave its consent, and the production is supervised. The public must know when AI is involved.

Voice-cloning tools are becoming accessible

The case of Sir Michael Parkinson shows what voice-cloning tools already make possible. Services such as ElevenLabs make these techniques easier to try than they were a few years ago. Open-source projects such as RVC, short for Retrieval-based Voice Conversion are also moving things in this direction.

This use can be sensitive, sometimes touching, when it keeps the voice of a deceased public figure alive with their loved ones' consent. The same technology can also fabricate false audio evidence. It can imitate a famous person without authorization. It can make a telephone scam much more credible.

The Touche pas à ma VF movement is campaigning specifically to protect voice actors from AI automation, with a petition dedicated to dubbing by humans for humans.

Verification also depends on context

With these synthetic voices, critical thinking must also consider the context. A voice can sound natural even though the content reveals that it was fabricated. If Sir Michael Parkinson mentions recent events that occurred after his death, the listener understands that the interview was produced afterward.

These clues may become more subtle. A better-written production with more carefully considered context would make detection harder. Transparency then depends on labeling, the source and the production conditions. The ear does some of the verification, but context becomes essential.

The BBC also addressed this subject in another article, Is there something special about the human voice?, devoted to human voices, audio deepfakes and the way AI blurs familiar reference points.

Today, with a little effort and the right tools, it is already possible to create a voice very close to the original. Virtually Parkinson demonstrates this in a case authorized by the family. The same level of realism, used without consent, poses a much more serious problem.

Yes, AI can copy a voice to the point where it is almost indistinguishable from the real one.