Is it allowed to transcribe a conversation under §201 StGB?

By Kyle Nelson, Founder, Fazit

By Fazit's founder. I'm not a lawyer.

In Germany, don't treat “Gespräch transkribieren” as automatically allowed under §201 StGB. The point is consent first: the source material says §201 StGB makes recording non-public speech without all-party permission a criminal issue, while German counsel reported that a RAM-only pipeline that transcribes in volatile memory and never writes an audio file generally does not meet the “recording on a sound carrier” element. This is not legal advice, and consent plus professional-conduct duties apply to the person running the call.

Gespräch transkribieren erlaubt §201 StGB

If you run confidential 1:1 client calls, your safest working assumption is simple: ask before you transcribe. That is true even if your tool says it doesn’t record, even if no bot joins, and even if the finished output is only a written note.

The hard part is that “transcribing” can mean three different technical acts. It can mean recording audio and sending it to a cloud speech service. It can mean streaming live audio to a vendor without keeping the audio file afterward. It can mean processing audio in RAM on your own Mac and writing only the resulting note. Those are not the same risk.

Fazit is built for the third path. A Core Audio process tap reads the call app’s output and the microphone directly. No bot joins Zoom, Meet, Slack or WhatsApp Desktop. The audio exists only in a fixed-size RAM ring buffer with no write API, and it is destroyed on every exit path including errors. Every note records audio_retained: false in its frontmatter.

The §201 StGB issue doesn’t disappear because a tool has a privacy-first architecture. The product can reduce the audio-file and vendor exposure, but the caller still has to handle consent before the conversation is captured.

What §201 StGB means for client-call notes

The primary legal source is §201 StGB itself, the German Criminal Code provision on the confidentiality of the spoken word. The source material for this post states the practical reading plainly: recording non-public speech without all-party permission is a criminal offense in Germany, with possible punishment up to 3 years.

That is the baseline for consultants, coaches, lawyers, financial advisers and therapists. If a client didn’t agree, don’t rely on a product page to save the workflow. Say what you’re doing before the call substance begins, and give the other person a real chance to refuse.

The nuance is the word “recording.” Fazit’s internal research says German counsel reported that a RAM-only transcription pipeline that never writes an audio file generally does not meet the “recording on a sound carrier” element of §201 StGB. That is why architecture matters: a tool that creates an audio file, uploads audio to a cloud, or stores a recording has created a different fact pattern from a tool that processes volatile memory and discards the audio.

That counsel note is not permission to skip consent. It is a reason to avoid creating an audio file in the first place.

Is transcription the same as recording under §201 StGB?

My opinion: for client work, arguing about labels is the wrong move. If the other person would reasonably want to know that their spoken words are being turned into a transcript or client note, tell them.

Still, the technical distinction is real. A conventional recorder creates a durable audio artifact. Many cloud notetakers also create transcripts, summaries and vendor-held records. Fazit’s design avoids the audio artifact: transcription runs on-device with Parakeet via CoreML, note generation runs through a local model on localhost, and the output is a plain Markdown file in your Obsidian vault or Apple Notes. There is no Fazit database holding a copy of the note.

That distinction is the same reason we write so much about architecture instead of feature checklists. If you want the deeper version, read the RAM-only architecture post and the on-device transcription post. The legal exposure starts to change when the audio file was never eligible to exist.

Here is the comparison that matters:

QuestionCloud or bot transcriptionFazit-style local transcription
Does a bot join the call?Often yes, depending on the toolNo
Where does call audio go?To a vendor cloud or third-party speech serviceStays on the Mac for transcription
Is call audio written to disk?Often yes, or held by the service during processingNo, it lives in a fixed-size RAM ring buffer
Where is the note stored?Often in a vendor databaseAs a Markdown file in your Obsidian vault, or Apple Notes
What remains after the call?Recording, transcript, summary or vendor-held record depending on the toolThe note and transcript text, with audio_retained: false

The strongest privacy line is not “we delete recordings later.” It is “the call audio is never written to disk.”

How to transcribe a client conversation without creating a recording

A good workflow has to be boring. Before the call, tell the client that you use a local note tool, that no bot will join, that audio won’t be kept, and that they can say no. Then start capture only after that notice.

On the Mac, Fazit captures two separate streams: your microphone as “You” and the call app’s audio as “Them.” That makes 1:1 attribution straightforward because the streams are physically separate. Fazit is deliberately scoped to 1:1 client calls. With 3 or more participants, remote voices are mixed into one system-audio channel, and separating them requires diarization, a quality and liability problem the product does not take on.

The processing path is short. The mic and call-app streams enter RAM. Parakeet via CoreML transcribes on the Mac. A local model on localhost writes the note. The app writes one Markdown file into your Obsidian vault. The audio buffer is reset.

There is a separate account lane. Models are downloaded once from public CDNs on first run, and account, licensing and payment use Fazit’s servers. That is separate from call content: call audio, transcripts and note content do not leave the device for note generation.

The artifact you keep is the note, not the audio.

Why no bot does not mean private

A botless product can still stream audio to a vendor. That is why “no bot joined” is not enough for lawyers, advisers, therapists, consultants or coaches with confidentiality clauses. The meeting can look clean to the client while the audio still leaves the machine.

This is also why consent law and confidentiality duties are different conversations. Consent asks whether the other person agreed to the capture. Confidentiality asks whether you disclosed client information to a third party. For lawyers, that raises privilege concerns, which we cover in the lawyer-focused article. For consultants, the NDA may already decide the answer, which we cover in the consultant article.

Fazit’s answer is architectural. No bot joins the call. Transcription runs on-device. Note generation runs locally. Audio is never written to disk. The finished Markdown note lands in a folder of your Obsidian vault, and there is no Fazit database holding a copy.

Botless describes the participant list. Private depends on the audio path.

What I would say before a German client call

Use plain language. Don’t hide behind tool names. A clean notice sounds like this:

“I use a local note tool to help me write the session notes. No bot joins the call, the audio isn’t kept, and the note is created on my Mac. Are you okay with that?”

That sentence does three jobs. It tells the client that speech will be processed. It tells them the audio won’t be retained. It asks for a yes before the call continues.

If your professional rules require more, use more. Lawyers, therapists, financial advisers and regulated professionals may have duties that go beyond general consent. Fazit does not hold SOC 2 or HIPAA certification, and nothing in the product removes your professional-conduct obligations.

The best compliance posture is the one you can explain in one sentence before the client starts talking.

What Fazit keeps after the call

Fazit keeps the note you asked it to write. The call audio is not stored, not uploaded, and not deleted later from a hidden recording store. It is never written to disk.

The note is a plain Markdown file in your Obsidian vault, or Apple Notes if you choose that output. It can contain a summary, notable moments, action items with owners and a follow-up email draft. The transcript text can appear as You: and Them: because 1:1 audio is captured as two separate streams.

That still means you have a written record. Written records can carry their own confidentiality, retention and discovery issues. If your profession has rules for client notes, apply those rules to the Markdown file. Fazit’s architecture removes the audio file from the problem, not the fact that you created a note.

The final durable artifact is the file in your vault.

Sources