Transcription for Mac · on-device
The recording never leaves your Mac.
VibeVerter transcribes audio and video on your own Mac — it labels each speaker, links every line to the second it was spoken, and saves the result as Word, PDF or plain text. No account. No upload. No monthly bill.
Apple silicon · macOS 14+ · free to try
That part already works. Every line links back to the audio — open the HTML copy next to the PDF and clicking a timestamp jumps the player straight to that moment.
Good. Then we ship it that way.
One window. Three jobs it does properly.
VibeVerter grew out of one person needing accurate transcripts of long calls without handing the audio to a website. Every part of it is built around that.
- Speakers
- Speaker detection separates the voices and labels them Speaker 1, Speaker 2, and so on. When the run finishes it plays you a few seconds of each and asks for real names; every line in the document updates at once. pyannote · community-1 model, bundled — no sign-up
- Timestamps
- Every paragraph is stamped with the time it was said. In the HTML copy the stamp is a link: click it and the built-in player jumps to that second. Word and PDF carry the same links back to the audio.
- Documents
- Export to Word (.docx), PDF or plain text — laid out as real speaker turns with punctuation and paragraphs, not one unbroken block. Hesitations and repeated false starts are tidied out; nothing is ever added.
- Convert
- Turn any audio file into MP3 at the bitrate you choose. Drag a whole folder in and it works through the queue, skipping anything that isn’t audio or video.
- Video
- Pull a clean audio track out of a video file — MP3, M4A, WAV, FLAC or OGG — without re‑encoding the picture or needing a separate tool.
- Engine
- Transcription runs on Whisper large-v3-turbo through Apple’s MLX framework, using the GPU in your M‑series chip. A one‑hour recording is done in a few minutes. first run downloads the model once (~1.5 GB); after that it is fully offline
A web transcription service keeps a copy of your recording on its servers. For a privileged call, a patient session, an unpublished source or anything under an NDA, that copy is the problem.
VibeVerter has no servers. The audio, the transcript and the model all stay on your Mac. Download the speech model once, then turn off Wi-Fi if you like — it still works.
That also means there is no per-minute meter and no seat licence. You are running a program on your own computer, the way software used to work.
Add, choose, start.
Add your files
Drag audio or video onto the window, or use Add Files / Add Folders. One file or a hundred.
Choose the task
Transcribe, convert to MP3, or extract audio. Pick the output format, the language, and whether to detect speakers.
Press Start
Watch the progress per file. When it needs the speaker names it asks, then writes everything to the folder you chose.
What it needs, what it speaks.
| Mac | Apple silicon — M1, M2, M3 or M4. Intel Macs are not supported (the engine is Apple-silicon only). |
|---|---|
| macOS | macOS 14 Sonoma or later. |
| Disk | About 2 GB free for the speech model, downloaded once on first use. |
| Internet | Needed only for that first download. Everything after runs offline. |
| Languages | Automatic detection, plus a fixed choice of: English, Arabic, Chinese, Dutch, French, German, Hindi, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Turkish. |
| Transcript out | Word (.docx), PDF, or plain text — each with an HTML copy that has the audio player and clickable timestamps. |
| Audio out | MP3, M4A, WAV, FLAC, OGG. |
Getting a copy
VibeVerter is signed and notarized by Apple, so it opens with no security warnings. It works immediately — transcription is capped at 10 minutes per file until you add a licence. This is an early release for friends, so licences are free for now. Just ask.
Free to try right now. No subscription, no per-minute charge.
Licences are free during this early-access period. Message me and I’ll send you a key.
Use it on the Macs you own. A licence key unlocks full-length files.
Built on open work
Speaker detection uses the pyannote speaker-diarization-community-1 model by Hervé Bredin and the pyannote team, released under CC BY 4.0 and redistributed inside VibeVerter under those terms. Transcription uses OpenAI’s Whisper model (MIT) in the MLX-community conversion, run through Apple’s MLX (MIT).
Media handling uses FFmpeg, LAME and the Xiph codecs, built and included under the LGPL 2.1. You may obtain the corresponding source for those components — including any changes — from vibeverter.app/licences, for no more than the cost of distribution. This offer is valid for three years from the date you received your copy.