AI Notetaker No Audio File Saved: What the Phrase Actually Means

10/20/2026

By Kyle Nelson, Founder, Fazit

ogTitle: "AI Notetaker No Audio File Saved: The Architecture Test"

An AI notetaker that saves no audio file is a claim about architecture, not a label. The question that matters is whether the tool creates a durable audio artifact at all, and where the sound lives before text exists. Fazit keeps call audio in a fixed-size RAM ring buffer with no write API, destroys it on every exit path including errors, and writes audio_retained: false into every note's frontmatter.

What "No Audio File Saved" Does Not Mean

The phrase gets used loosely. Granola's documentation says its macOS and Windows applications transcribe meetings in real time and do not store meeting audio, retaining the transcript instead. That is a retention policy, not an absence of an audio file. The audio still exists during the call, still travels to Granola's systems and third-party transcription services, and is deleted afterward. The file existed. It was saved somewhere, even if briefly.

Wispr Flow's documentation is more specific. It says uploaded meeting audio is deleted seven days after upload, and local meeting audio is deleted seven days after the recording was last written. A recording interrupted by an application quit or restart is finalized the next time the application opens. That means a local copy sits on the computer to support resumption or retry. The audio file exists on disk for up to a week. "No audio file saved" is not what Wispr Flow claims; it claims a deletion schedule.

The distinction matters for confidential client calls. A consultant's NDA often restricts disclosure to third parties. A lawyer's privilege analysis starts with where the client-call audio, transcript, prompts, logs, and summaries go. A therapist's session note is governed by consent and professional-conduct duties. In each case, the question is not whether the vendor eventually deletes the audio. The question is whether the audio ever becomes a file that could be intercepted, subpoenaed, or disclosed before deletion.

The Architecture Test

You can verify the claim without reading a security page. Ask five questions about the pipeline:

Does a bot join the call as a participant?
Does the audio leave the Mac for transcription?
Is the audio written to disk at any point, even as a cache?
Does the vendor hold a copy of the transcript or note after the call?
Can the user delete everything the tool knows?

Fazit's answers are specific. No bot joins. A Core Audio process tap reads the call app's output and the microphone directly. Transcription runs on-device using Parakeet via CoreML, and note generation runs against a local model on localhost. The audio exists only in a fixed-size RAM ring buffer with no write API. The output is a plain Markdown file written into a folder of your Obsidian vault, or Apple Notes. There is no Fazit database holding a copy.

The README in Fazit's repository states the invariant directly: audio exists in exactly one place, AudioRingBuffer, and is destroyed in exactly one place, CaptureController.destroyAudio(), which runs on every path including errors. There is deliberately no write(to:) on the buffer.

Why the Audio Path Decides the Risk

The Otter.ai wiretap class action gave the category a case number and a price: $5,000 per recording under California Penal Code § 637.2, plus a transcript that becomes discoverable evidence in any lawsuit. The Brewer v. Otter.ai case file walks through how one February 2025 sales call became four consolidated class actions. The complaint alleges the bot recorded meetings without everyone's consent and fed audio into model training. No court has ruled yet, but the complaint is a checklist of what any cloud notetaker does with your calls.

The Granola lawsuit raises a different angle. A proposed class action filed on 30 July 2026 in the Northern District of California argues that the thing everyone liked about Granola, that no bot appears in the call, is the same thing that left other participants with no way to know a recording was running. Bot-free capture removes the visible signal. It does not remove the recording.

For lawyers, the risk is consent and privilege together. Sending client-call audio to a vendor and its subprocessors is a third-party disclosure that invites waiver arguments and creates a discoverable record. The attorney-client privilege analysis explains why on-device capture avoids the disclosure, and the client-notice duty it does not.

For accountants, the stakes are statutory. IRC § 7216 makes unauthorized disclosure of tax return information a misdemeanor, and § 6713 adds a civil penalty per disclosure. A cloud notetaker sitting on a client call is a disclosure question. The § 7216 analysis details what that means for a tax return preparer.

What the Comparison Table Actually Shows

The tools everyone shortlists all write a good note now. The comparison that survives a security review is who holds a copy of your conversation after the call ends.

Tool	Bot joins	Audio leaves device	Audio file on disk	Vendor holds transcript
Fathom	Yes	Yes	Yes	Yes
Otter	Yes	Yes	Yes	Yes
Fireflies	Yes	Yes	Yes	Yes
tl;dv	Yes	Yes	Yes	Yes
Granola	No	Yes	Yes, then deleted	Yes
Fazit	No	No	No	No

Granola is the best of the bot-free notetakers, but it still streams audio to third-party transcription services. The Granola subprocessors post names OpenAI and Anthropic, US-only AWS storage, no EU residency, and transcripts retained indefinitely by default. The audio reaches Granola's systems and third-party services before it is deleted.

Fazit is the sixth row, where the copy was never eligible to exist. The architecture comparison makes the point: five of six create a durable artifact on someone else's infrastructure. Fazit is the pipeline where the artifact never exists.

The German Angle

In Germany, §201 StGB targets recording non-public speech without all-party permission. It is a criminal offense, up to three years. German counsel states that a RAM-only pipeline that transcribes in volatile memory and never writes an audio file generally does not meet the elements of §201 StGB, because there is no recording on a sound carrier. The §201 StGB analysis treats client-call transcription as a consent question first.

This is not legal advice. Consent and professional-conduct duties apply to the person running the call, regardless of the tool's architecture. Fazit's architecture reduces the vendor's and the audio-file exposure. It does not remove your duty to have consent where the law requires it.

The One Artifact You Keep

A note is a record you control. A recording is a record you must defend. The financial advisor analysis puts it plainly: no SEC or FINRA rule makes an advisor record client calls, but plenty of rules punish records you can't produce. Since June 3, 2026, the amended Reg S-P covers every firm's vendor cloud. The compliant posture is one artifact: a note you retain under your own policy, with no recording behind it.

Fazit writes that note as a plain Markdown file into your Obsidian vault. The frontmatter records audio_retained: false. The audio was never written to disk, never sent to a vendor, and never eligible to be subpoenaed. The note is yours to grep, back up, and delete.