Episode 7 · Article on the episode

Workshop report on podcast production: where AI helps and where it is only a tool

A podcast about AI that is produced almost entirely by hand. The workshop report shows where automation genuinely carries the work and where it is merely another tool in the box.

By Mark Zimmermann · 15 Sep 2025 · 4 min read · Auf Deutsch lesen

This solo episode is a workshop report rather than a topic episode. Which tools actually keep the production running, and where does AI really help.

The most honest statement comes at the end: the podcast itself is produced completely by hand. Generated voices appear only in the intro.

Editing by text

The most interesting move is the edit. Riverside displays the spoken material as a transcript per speaker, so slips of the tongue can be removed by marking them in the text instead of hunting for them in the waveform.

That is a good example of an improvement that brings no new capability but moves a familiar activity onto a more suitable tool. Speech is text, and text can be read and marked. A waveform has to be listened to.

Added to that is voice cloning for the case where a word or a whole sentence has to be generated afresh. This is the point where an editorial boundary runs: a removed slip of the tongue changes nothing about the statement. A subsequently generated sentence that nobody said does change it.

The audio processing

Auphonic handles noise reduction, pause shortening and loudness levelling, and performs better at it than Adobe Podcast, which struggles above all with breathing sounds and unnecessary pauses.

For guest episodes the procedure becomes more elaborate, and the order of steps is the actual trick.

Why multi-track processing runs in two stages

Every audio track runs through the processing individually. That is necessary because noise reduction and loudness levelling have to respond to each recording situation on its own terms: one guest in a home office, one in a studio, one on a notebook microphone.

The tracks are then brought back together in sync via Ferrite Recording Studio.

Afterwards the combined track runs through the processing a second time, this time exclusively to shorten pauses. The reason: a pause is only a pause when nobody is speaking. On an individual track, every passage where the others are talking looks like a pause.

Anyone who does this second pass on the individual track cuts away exactly the passages where the conversation partners are speaking, and destroys the conversation.

Once the edit is finished, everything lands at Podigee: title, subtitle, show notes and cover are maintained, and automatic transcription is switched on. For guest episodes there is an approval loop before publication.

The numbers

What is remarkable about this episode is the willingness to name numbers. Download figures for individual episodes are given, along with a total heading towards 750 downloads.

That is not a reach success and it is not sold as one. It is a position fix, and it is more useful than any success announcement: anyone building something themselves gets an honest order of magnitude here for where a specialist podcast stands after seven episodes.

Marketing by agent

With marketing it gets concrete. Alongside classic posts on LinkedIn, an agent reads out the show's own transcript, extracts keywords and searches for matching posts on social networks, in order to leave appreciative comments there with a link to the episode.

That is a usable example of content exploitation without a marketing department, and it calls for a boundary that resonates through the episode. A comment under someone else's post is an intervention in someone else's conversation. It carries weight when it fits the subject and has obviously been read. It does damage when it looks like a keyword match.

In practice that means automation up to the suggestion, approval by a human. The effort per comment then comes to ten seconds, and the difference in the result is considerable.

Conclusion

The workshop report is valuable because it shows the boundary at which automation stops being sensible.

Everything that is a manual step gets automated: noise, levels, pauses, transcription, chapter marks. Everything that demands judgement stays by hand: what stays in, what has to go, which sentence carries the statement.

For transferring this to your own projects, the order of steps is instructive. First automate the manual steps whose results you can judge yourself. And if you take one thing from this episode, take the two-stage approach for multi-track recordings. It costs one extra pass and rescues the conversation.

To close, a look at further podcast experiments automated to varying degrees: the comedy show “404 Lachen nicht gefunden”, the learning journey “Kopf und KI”, the fictional criminal cases of “Schattenakte”, the prompting format “Prompt Intelligence” and the English-language “AI Revolution”.