Meeting transcription that never leaves your Mac.

Who needs local transcription

Some calls cannot be uploaded to a transcription service: attorney-client conversations, patient sessions, source interviews, M&A discussions, anything under NDA. The moment audio hits a vendor’s cloud you inherit their retention policy, their subprocessors, their breach surface, and a data-processing agreement to negotiate. Local transcription removes that whole category of problem: there is no vendor in the loop, so there is nothing to agree, audit, or trust.

What exactly “on-device” means in echo99

• Recording: your microphone and your Mac’s audio output are captured as two separate WAV files, straight to local disk.
• Speech-to-text: recognition models are downloaded once to your Mac and run locally. After that download, everything works with Wi-Fi off.
• Speaker labeling: diarization, which works out how many people are speaking and which words belong to whom, also runs locally. Your mic track is “Me”, other voices become “Them-1”, “Them-2”, …
• Networking: echo99 ships with limited telemetry, meaning a launch-time update check that sends an anonymous install id, and nothing else. Audio and transcripts are never uploaded, and there is no account to create.

Your data lands in ~/Documents/echo99/ as plain WAV and JSON files, one folder per recording (you can pick another location in Config → Storage). Deleting a recording’s folder removes its audio, transcript, and notes. If speaker recognition is on, the voices you name also live in a gallery on your Mac: remove someone in Manage Voices, and choose Delete and remove from past recordings to erase their speaker data from your recordings too. Plain local files help when your work is governed by data-minimization rules, though whether your recording practice satisfies any given regulation is a question for your counsel, not an app.

What you get after each call

• transcript.txt and transcript.json: the full conversation with speaker labels and timestamps, readable and machine-parseable.
• mic.wav + system.wav: the clean source audio, one track per side of the conversation, kept separate on purpose.
• metadata.json: when it was recorded and with what settings.
• summary.md, once you summarize the recording: notes written by a language model that also runs on your Mac.

Limits

• The transcription models download once from the internet (a few hundred megabytes); after that, echo99 is fully offline-capable.
• On-device recognition is strong, but heavy accents, crosstalk, and bad audio still produce errors, as they do in any transcription, cloud or local. A custom-vocabulary option helps with names and jargon.
• The default speech model handles 25 European languages, Czech included; the technology page lists them and the English-only alternatives.
• Apple silicon transcribes fastest; Intel Macs are supported but noticeably slower.

New to echo99? See getting started for setup and your first recording.

Related guides: Transcribe interviews, Record system audio, Transcribe audio and video files