I Cloned My Voice. And I Tell Everyone.

From idea to podcast in two days, and why transparency is the question that actually matters.

---

On Wednesday evening I had the idea. By Friday afternoon the podcast was live on Spotify and Apple Podcasts.

In between came a voice clone, an AI host named Matilda, and 17 briefings from my newsletter that you can now listen to instead of read.

Season 1 is called “Rewind.” It tells my RAMmageddon briefings backwards: from the latest part, “The Mediator Between Worlds,” back to RAMmageddon part 1 from February 2026. Matilda hosts, I speak as an AI clone. The whole thing exists in English and in German, together around seven hours of audio.

Here is how it went. And why I believe the question that actually matters is a different one.

The first attempt sounded like I was behind a wall

I started locally. An open-source model on my Mac mini, everything on my own hardware. The result was muffled, as if I were standing behind a wall. Fine for an experiment, not for seven hours of listening.

So, ElevenLabs.

Two minutes against fifty minutes

For the German edition, an instant clone from a two-minute voice memo was enough.

For the English edition I trained a professional clone. For that I recorded just under 50 minutes of English, using the Voice Memos app on my iPhone. No studio, no microphone setup, no voice actor.

The biggest construction site was pronunciation

The problem was not the voice itself but individual words. “RAMmageddon.” “DRAM.” And my own name: “Yens,” not “Dschens.” “Klasen” with a long A.

I solved it through adapted spellings in the script and a more stable setting for the host’s voice. Sounds trivial. But it accounted for the largest part of the work.

The AI found errors in my own texts

While creating the scripts, something happened that I had not expected: the AI found inconsistencies in my original texts. Passages nobody had noticed while reading, but which no longer held together when spoken.

I made the decisions. Every change, every phrasing. The AI delivered, I took responsibility.

What it costs

An ElevenLabs subscription in the double-digit to low triple-digit dollar range. No studio. No voice actor.

And the two days between idea and publication were not two working days. I worked on it after hours, whenever a bit of time opened up between work and family. Anyone who sits down for it deliberately will, by my estimate, be done in less than two hours including the voice recordings.

This is the point where you have to be honest: the technology is here. For everyone.

Not perfect, but a start

Honestly: the result is not perfect yet.

My wife was impressed. My children noticed immediately. Colleagues and friends say it is very impressive, but that in some places you can still hear it.

That matches my own impression. And I know why. I made my recordings on the side, not in the quietest room and with considerably less material than would have been possible. For the professional clone you can upload up to two hours of voice recordings. That is exactly what I will do next, in a quieter room and with more material. Then I will see how the next episodes sound.

For a first shot after such a short time I am satisfied nonetheless. And that is precisely the part that occupies me most.

Two minutes of voice memo. One subscription. Less than two hours if you sit down for it deliberately. That is all it takes to produce a voice that impressed even my wife.

What I did with my own voice, anyone can do with someone else’s. With a voice message from the family chat, a clip from a webinar, a video on LinkedIn. The call from the supposed boss who urgently needs a transfer. The voice message from the supposed child who is in trouble. The technology for that does not sit in a lab, it is one click away.

That is no reason to keep your hands off it. But it is a reason to look closely at how you use it.

The question is no longer whether, but how honestly

The transparency obligations of the AI Act have applied since 2 August. For audio, the draft of the Commission guidelines names a spoken notice at the beginning as a form of labelling. Spotify now issues verification badges and intends to remove podcasts that imitate other people’s voices without consent. According to Bloomberg, in a single week in April around 39 percent of all new podcasts were AI-generated.

That shows: it is not about whether AI voices are coming. They have long been here. It is about how openly you deal with them.

Cloning my own voice is not something I find objectionable. It is my voice, my consent, my content. It becomes objectionable when listeners are meant to believe they are hearing a human when they are not.

That is why every episode says at the start that the voices are AI-generated. And wherever there are connections to my employer, I disclose them. For a private project that would probably not even be mandatory. I do it anyway, because trust is created or lost at exactly this point.

Listen in

English
Spotify: https://open.spotify.com/show/0styyqtXSQk2Ipxr8apxvS
Apple Podcasts: https://podcasts.apple.com/de/podcast/id6818447410

German
Spotify: https://open.spotify.com/show/6TO00S4HGaBNaSr5OKodtq
Apple Podcasts: https://podcasts.apple.com/de/podcast/id6818628457

Both editions
YouTube: https://www.youtube.com/@JensKlasenAI/podcasts
Website: https://klasen.ai/#podcast

Transparency: The voices in the podcast are AI-generated. Matilda is an AI voice from the ElevenLabs library, my voice is an AI clone of my real one. The cover image was created with Gemini. The ideas, analyses and arguments are mine. Since 16 September 2026 I have been a Cloud Native Architect at Everpure (formerly Pure Storage), on the Portworx team. The podcast is a private project, the opinions are my own.

Jens Klasen writes in the Future Proof Tech Briefing about the upheaval in enterprise IT. You can find all previous installments in my newsletter and at klasen.ai.