The idea
The journal is text. But many people would rather listen than read: in the car, on a run, on the side. So the essays should also appear as audio. The test ran on the latest essay, “I’m dangerous — so buy me”: about makers who warn against their own AI. That machines should voice this particular text has its own irony.
It started with a reading and a conversation between two AI hosts in the style of NotebookLM. The conversation sounded creepy; the reading was correct, but clearly read out. The reading became a feature. This report is about how that feature developed.
The format
My voice narrates and delivers the figures. A second voice steps into the scenes in the present tense: the TV network in 1988, the cinema in 1973, the Senate hearing, the record store, the test lab, the Council of Elrond. Ambience runs under the scenes, music beds under the narration. An opener gives the title and the AI notice; closing credits carry the editorial notes. The running text has a narrator, not an author reading himself aloud.
Nine versions
| Version | What changed | Trigger |
|---|---|---|
| v1 | The reading becomes a feature: scene voice, ambience, music | “alternating voices” |
| v2 | Music beds that follow each chapter | “editorial music under the narration” |
| v3 | English titles in English, Margenfeld with a hard g, new mix chain | “Private Parts is pronounced in English” |
| v4 | Breaths kept, clicks repaired, clone vs. Ben | “cut-off breaths and clicks” |
| v5 | Opener, narrator stance without “I”, closing credits | “classic narrator stance” |
| v6 | Eleven v4 through the connector, direction with tags | my own test of v4 |
| v7 | Music without piano, more drive | “lose the tinkling piano” |
| v8 | Audiobook tags, softer mix chain | “less pressure, more audiobook” |
| v9 | Voice-over without any tags, music from Suno | “that’s where it gets creepy” |
Three models, one voice
The same Listerine passage in five versions (German). Up to v5, ElevenLabs Multilingual v2 speaks; from v6 on, Eleven v4. The scenes are still voiced with v3.
The difference shows less in the long paragraph than in the short sentences. That’s where the audiobook style tips over:
What the tags do
Eleven v4 can be directed with tags in square brackets. They work, just not the way you’d expect.
- [curious] raises the voice. On “Tolkien hat diese Figur zu Ende gedacht” the pitch rose to 188 instead of a median of 143 hertz.
- [softly] and [calm audiobook narration] make short sentences breathy. In a long paragraph they lower the pitch by 1.6 semitones and take out pressure. In short sentences the melody goes flat and the pauses grow. That’s exactly where I heard “creepy”.
- Without tags, my clone is at its most even. The mix takes the pressure out instead: one gentle compressor instead of two.
- Pronunciation belongs in phonetic script, not in respellings. v4 understands IPA between slashes. My name is written as /ˈmaʁɡn̩fɛlt/ and comes out with a hard g in every take.
The music
The first beds came from ElevenLabs Music. They were too flat, and the piano at the start grated. That led to a dedicated set: ten music beds as a suite with one shared palette of double bass, cello, a dry pulse and a clean guitar. No piano, no vocals, no genre clichés. I found the first round of prompts too cheesy; I generated the second in Suno. All ten beds sit in the asset library with prompt, measurements and description, and from now on they carry every episode. The intro and outro come from the signature, MF 01.
The English version
The essay already exists in English on margenfeld.com. That text was switched to a narrator stance and given an opener and closing credits. My clone narrates in English too, without tags. The scenes are voiced by Heather Faith, a calm American voice from the library, “strong without theatrics”. Ambience, beds and mix are the same. My clone speaks short sentences a little higher in English; three variants changed that only slightly.
Why this version holds up
The feature can’t match a real production. It remains a voice from the machine. Still, v9 is acceptable, and for reasons that can be stated:
- The text is right. Every line is checked against the script by speech recognition, in both languages.
- The pronunciation is right. Names and titles are checked by phoneme recognition, not guessed by ear.
- The voice stays in its lane. No more drama patterns: no block drops below 8.5 semitones of range, and pauses stay within the usual range. Where a short sentence came out too high, it was regenerated.
- The mix meets the standard. −16 LUFS, true peak below −1 dB, measured after encoding, clicks repaired.
- And the ear signed off: “That works.”
What remains: the transitions between sections could be smoother, and the short English sentences sound a touch too bright.
A look at the case
Hardly anyone will hear a single breath or click. But listeners feel the sum. An audio piece that breathes slightly wrong in twenty places doesn’t sound broken. It sounds artificial. The radio ear is no obstacle here. It is the measuring instrument the machine does not yet have.
Nine rounds turned the experiment into a method: a narrator without tags, phonetic script for names, a fixed set of music beds, a mix chain with a standard, and a check that hears every line before I do. The next episode doesn’t start from scratch.
What this means for an organization
Strategy. The first case is an experiment. Value comes with the second, when it does not start from zero again. After nine versions there is a narrator without tags, phonetic spelling for names, ten music beds and a mix standard. A good hour and a half of my time went into the set of beds alone. That is an investment in every further episode.
Role. Someone keeps the standards: −16 LUFS, the phonetic spelling for names, the beds in the asset library with prompt and measurements. Without that role, a method falls apart into single cases again.
Process. Every version is recorded with what triggered it. “Lose the tinkling piano” sits next to v7. So the next round knows why something is the way it is and does not have to find out again. An organization that files only results forgets exactly that.
What stays human. The ear. The check listens to every line before I do. Whether the whole sounds artificial is still decided by a human: “That works.”