Last week I tried to build a German-language AI podcast from a single PDF source. No budget, no studio, no microphone. Here is the workflow that actually worked in May 2026.
The 45-minute workflow
The goal was a 15-minute episode summarising a research PDF in conversational German. The full setup:
- Drop the PDF into NotebookLM and switch language to German. The new German voice model released in early 2026 is shockingly natural, no English accent.
- Generate the Audio Overview. Three minutes later you have two AI hosts having a real conversation about the document.
- Run the output through Adobe Podcast Enhance if you want to tidy up the final mix. Free tier covers one hour per month.
- Add intro and outro music from Pixabay Music. KI-generated music from Suno or Udio is NOT licensed for commercial use on the free tier - I learned this the painful way.
- Upload to Spotify for Podcasters. Free.
Total cost: zero euros. Total time: 45 minutes.
The catch nobody mentions
Since August 2026 the EU AI Act Article 50 requires clear labelling of synthetic audio. One sentence in the intro plus a note in the episode description satisfies the rule. If you clone a real person's voice, you also need their written consent under GDPR Article 9 (biometric data). Skipping this is a fast track to a cease and desist letter.
What the German-language tooling actually looks like
ElevenLabs Multilingual v2 still produces the best German voice clones. KikiVoice from Frankfurt is the only fully GDPR-compliant alternative I could find, with EU hosting. Adobe Podcast in German is good but slightly synthetic on long passages.
For a full German-language breakdown with 7 tools compared and a Shopify-style decision table, see this comparison on KI Tools DE.
What I would do differently next time
I tried to skip the music step. Do not. A podcast without intro music sounds amateur in the first three seconds and loses 60 percent of listeners. Pixabay has thousands of tracks that are properly licensed and free to use.
The other lesson: NotebookLM is brilliant at conversational explainers but bad at lists. If your source PDF is a checklist or a table, the AI hosts will skip half of it. Convert the source to prose first.
Happy podcasting.
Top comments (0)