Does transcription store audio in 2026?

09/17/2026

By Kyle Nelson, Founder, Fazit

By Fazit's founder, not a lawyer.

Transcription doesn’t have to store audio. Many transcription tools keep audio for a period, delete it after transcription, or retain it as the primary record, so the right question is where the audio exists after the words become text. Fazit’s answer is stricter: for 1:1 client calls on macOS, call audio is read into RAM, transcribed on the Mac, and never written to disk.

does transcription store audio?

Sometimes, yes. A transcription product can store the raw audio, store only the transcript, store both, or avoid creating a stored audio file at all. Those are different architectures, and they create different obligations for a consultant after a strategy call, a lawyer after an intake call, or a therapist after a session.

According to the AEPD, Spain’s data protection authority, voice transcription services must apply data protection rules to metadata and to communication content when those data relate to a natural person. That matters because a client’s spoken name, health detail, tax detail, legal problem, or investment goal can sit inside the audio before it ever becomes a transcript.

The simplest mental model is this: audio is the source artifact, the transcript is the derived artifact, and the note is the working artifact. If your tool stores all three, your retention and disclosure problem has three copies. If it stores only the transcript and note, there are two. If the audio never becomes a file, the audio copy is absent from the record you manage.

For a coach who ends a Zoom call and sends a follow-up email, the useful output is the note, the action list, and the email draft, since the raw waveform rarely helps the client or the coach after the session ends.

where transcription audio is stored

The storage location decides who can touch the call content. Cloud transcription sends audio to a vendor. On-device transcription keeps the speech recognition step on the user’s machine. Fazit uses the second path for call content: transcription runs on-device using Parakeet via CoreML, and note generation runs through a local model on localhost.

The site’s architecture statement says Fazit captures the call app’s output and the microphone directly through a Core Audio process tap. No bot joins the call, and no participant sees a meeting assistant enter the room. The same statement says audio exists only in a fixed-size RAM ring buffer with no write API, and the buffer is destroyed on every exit path including errors.

That distinction matters during an ordinary client call. A financial adviser may need a written note of what the client asked and what follow-up was promised. The adviser doesn’t automatically need a replayable audio file of the client’s voice sitting in a vendor account.

For more on why local speech recognition changed the privacy posture of this category, see cloud transcription was a hardware workaround. The engineering point is narrow: on a modern Mac, the speech model can run close enough to the conversation that the audio doesn’t need a server round trip.

stored audio versus deleted audio versus RAM-only audio

This comparison is the point. “We delete audio” and “we never write audio” sound similar in a sales demo, yet they describe different facts in the pipeline.

Transcription design	What happens to the audio	What remains after the call	Example from a working day
Stored audio	The recording is kept as a durable file or object	Audio, transcript, notes	A lawyer can replay an intake call later, and that recording now exists as a retained client artifact
Delete-after-transcription audio	The audio is stored temporarily, then removed under a policy	Transcript, notes, plus deletion logs or vendor records	A consultant gets a summary, then relies on the vendor’s deletion period and storage controls
RAM-only audio	The audio is processed in memory and never written to disk	Transcript and note only, if the product writes those	A coach finishes a 1:1 call and keeps a Markdown session note in their own vault

The GDPR Advisor article on voice-to-text services says a common minimization strategy is to retain transcripts while deleting raw audio recordings once they are no longer needed. That is a reasonable retention pattern for many systems, but it still assumes the raw audio existed as stored data first.

A practical retention checklist from SayToWords recommends deleting audio after transcription if it is not needed and setting clear retention periods. The voice-dictation enterprise security guide from Weesper Neon Flow recommends defining retention periods and automatically deleting voice recordings after transcription, or within 30 to 90 days maximum. Fast.io’s transcription workflow storage guide gives a different operational pattern: keep audio for 90 days, keep transcripts for 2 years, and delete processing artifacts immediately.

Those examples show the range of normal practice. Fazit picks a narrower path for 1:1 client calls: the stored artifact is the Markdown note in the user’s Obsidian vault, and each note records audio_retained: false in its frontmatter.

how to tell whether a transcription tool stores audio

Ask where the audio exists at each step. Don’t stop at “private,” “secure,” or “bot-free,” because those words don’t answer the storage question.

A useful review takes five minutes:

Does a bot join the meeting as a participant?
Does the product send call audio to a vendor cloud for transcription?
Does the product create an audio file before transcription starts?
Does the product keep the raw audio after transcription?
Does the product store the transcript or note in a vendor database?
Can you export or delete every retained artifact yourself?

That list helps in a normal client workflow. If a consultant signs an NDA that restricts third-party disclosure, the audio path matters before the note template does. If a therapist or coach discusses sensitive personal facts, the existence of a stored voice file changes the practical risk. If a lawyer is worried about privilege, a third-party transcription path creates a separate question from note quality.

Fazit’s public architecture answer is specific: call content stays on the Mac for transcription and note generation, while account, licensing, payment, updates, and optional product analytics are separate from call content. The site also says Fazit makes no SOC 2 or HIPAA certification claim.

For a broader comparison of this exact question, see where does the audio actually go. For the separate issue of visible meeting bots, see an AI notetaker should not be a participant in your call.

does transcription create a legal recording?

Architecture can reduce the number of artifacts, but it doesn’t erase consent duties. Nothing here is legal advice, and consent and professional-conduct duties apply to the person running the call.

The AEPD article says GDPR applies to voice-transcription service metadata and to the content of the communication when those data relate to a natural person. That is enough reason for European professionals to treat voice data as client data from the first second of capture. The product choice then becomes a data-minimization choice: create fewer durable copies, store them in fewer places, and keep the artifact you actually need for the work.

In the United States, consent law is a separate question from storage architecture. Fazit’s existing post on whether it is legal to record client calls covers the consent map and explains why the recording you create is a second risk. The point for this article is simpler: if your transcription tool stores raw audio, you have a retained recording to manage under your own duties.

Germany has its own issue under §201 StGB, and Fazit’s German-focused article on whether it is allowed to transcribe a conversation treats client-call transcription as a consent question first. The safer working habit is to tell the other person what is happening before the call proceeds.

A plain client sentence does the job better than a hidden setting: “I use a local tool to turn this call into notes, the audio isn’t kept, and you can say no.”

what Fazit stores after transcription

Fazit writes one ordinary file: a Markdown note in a folder of your Obsidian vault, or Apple Notes if you choose that output. There is no Fazit database holding a copy of the note.

The repository README describes the 1:1 pipeline as two physically separate audio streams: your microphone is “You,” and the call app’s audio is “Them.” That gives clean You and Them attribution for 1:1 calls because the two sides arrive as separate streams. The README also says the app is deliberately scoped away from multi-party speaker labels, mobile capture, CRM sync, always-on auto-detection, and thread summaries.

That scope fits the working day of a consultant or coach. After the call, you need the summary, the moments worth remembering, the action items with owners, and a paste-ready follow-up email. You don’t need a retained audio file to ask, “What did Sarah agree to send before Friday?”

Fazit’s source statement also makes a narrow claim about “nothing leaves the Mac”: it applies to call content. Models are downloaded once from public CDNs on first run, and account, licensing, and payment use Fazit servers through a separate lane from call content.

The storage answer is therefore exact: Fazit stores the finished note where you asked it to store the note, and it doesn’t store the call audio.

what to ask before your next client call

Ask the vendor to describe the audio path in one paragraph. If the answer skips the words “audio file,” “transcription,” “retention,” and “deletion,” ask again.

For a lawyer, the question might be: “Does any third party receive the client’s voice?” For an adviser, it might be: “Will this create a recording I must preserve or produce under my firm’s policy?” For a therapist or coach, it might be: “Can I explain this honestly to the client before the session starts?”

The answer you want is factual enough to survive a security review. It should name where the audio is captured, where transcription runs, whether raw audio is written to disk, whether any vendor receives call content, where the transcript is stored, and how deletion works.

Fazit’s answer is short because the pipeline is short: Core Audio reads the mic and call app, RAM holds the audio long enough for local transcription, a local model writes the note, and one Markdown file appears in your vault.