Private Transcription Tools: Hosted vs Local
Which transcription options keep a recording off third-party clouds: hosted services, Whisper apps like MacWhisper and Buzz, and self-hosting.
The Scriptorium team ·
If you are transcribing a client call, a patient interview, a legal deposition or an internal strategy meeting, "is it free and accurate?" is the second question. The first is where the audio goes. The answer divides the market into three groups: hosted services, where the file is uploaded to the vendor's servers; local apps, where a speech model runs on your own computer; and self-hosted servers, where the same kind of model runs on a machine you or your organisation control.
We make software in the third group (and a hosted site as well), so we say where our own hosted site belongs: in the first. The two other groups below each have real strengths and we try to describe them fairly.
Group 1: hosted services
Otter, Notta, TurboScribe, Riverside's tool, VEED, Sonix, Happy Scribe, Granola and Tactiq are all, in different ways, services: audio or captions go to the vendor, and your transcript lives in their account system. They can be entirely appropriate. The questions to ask are on their own privacy pages: how long is the audio kept, who can access it, and is it used to improve models. We checked a few details on 2026-10-21:
- Granola says it uses your computer's audio and invites no bot, but its page does not say where the audio is processed.
- Tactiq says it works with no bot, through a Chrome extension for Google Meet, Zoom and Microsoft Teams.
- Our own hosted site, tryscriptorium.com, is a hosted service too: a no-account preview file is deleted after 24 hours, and an account keeps its files until you delete them (see our privacy page). Our self-hosting page says plainly that the hosted workspace suits non-sensitive files.
A hosted service is the right answer when convenience matters, the content is not confidential, and the vendor's terms are acceptable. It is the wrong answer when a contract or regulation forbids third-party processing, whatever the privacy page promises.
Group 2: local apps on your own computer
A local app downloads a speech model once and then transcribes on your machine. Two well known ones, both built on OpenAI's open Whisper model family:
MacWhisper
MacWhisper is a Mac app from Good Snooze. Its page lists a free tier ("Free forever") and a one-time Pro purchase at €64 with lifetime updates. It says sensitive files can be handled "without data ever leaving your Mac", supports 100 languages, exports subtitle and text files (SRT, VTT, TXT), and offers automatic speaker recognition. Local is not the only mode: the page also mentions connecting AI providers with your own API keys, so local-only is a setting you must keep, not a guarantee of the app. It is Mac-first, and its page does not say which features are in the free tier.
Buzz
Buzz is open source (MIT licence), described in its README as "Transcribe and translate audio offline on your personal computer. Powered by OpenAI's Whisper." It runs on macOS (Apple silicon only in current versions), Windows and Linux, exports TXT, SRT and VTT, and has no price. The README notes that its Windows build is unsigned, so Windows shows a warning during install.
Both are excellent for one person on one machine. The limits are structural: your computer does the work, so a long file on a slow laptop takes a while; there is no shared library for a team; and keeping files secure afterwards is up to you. For one confidential interview a week, this is often the best answer and it costs little or nothing.
A note on the model itself
OpenAI's Whisper README states that code and model weights are released under the MIT Licence, that there are six model sizes, and that "performance varies widely depending on the language". Smaller models are faster and need less memory; larger ones need a good GPU or patience. This is why accuracy for the same recording differs between a laptop running a small model and a server running a large one, and why you should test with your own audio. See our local Whisper guide for the practical setup.
Group 3: self-hosted servers
Self-hosting means running the application on a machine your organisation controls, so several people can use one private installation. Scriptorium can be deployed this way: the recordings, transcripts, search index and exports stay on the infrastructure you administer, the speech model runs there, and no third-party transcription API is required (our self-hosted page describes the deployment). The software offers 33 languages plus automatic detection and exports TXT, timestamped TXT, SRT, VTT, Markdown, JSON and CSV. The cost is work: someone has to install it, keep it updated, set up HTTPS and back it up. If nobody in your organisation can do that, a local app is simpler. See self-hosted vs cloud for the trade-offs in more depth.
Comparison
| Hosted service | Local app (MacWhisper, Buzz) | Self-hosted server | |
|---|---|---|---|
| Where audio goes | Vendor's servers | Stays on your computer (check settings) | Stays on your server |
| Setup | None | Install an app | Deploy and maintain a server |
| Team sharing | Built in | Not really | Yes |
| Cost | Free tier, then subscription | Free (Buzz) or one-time (MacWhisper Pro, €64) | Your hardware and time |
| Speed | Vendor's hardware | Your computer | Your server |
| Updates and backups | Vendor | You (app updates) | You |
Checked on 2026-10-21. MacWhisper, Buzz, OpenAI Whisper, Granola and Tactiq: their own pages. Prices and limits change; confirm on the vendor's page before you buy.
What none of them solve
- Consent to record. Where the audio is processed does not change whether you may record it.
- Retention on your side. A local transcript in a synced folder is a copy in someone's cloud. Know where your own files go.
- Accuracy. The same model can run anywhere, so privacy does not buy accuracy and accuracy does not need a cloud. See how accurate AI transcription is.
How to decide
- Does a contract, law or client forbid uploading the recording? Then group 2 or 3, no exceptions.
- Is it one person, occasionally? A local app.
- Is it a team, or recurring? Look at self-hosting, or at a vendor whose terms your organisation has reviewed.
- Is it ordinary, non-sensitive audio and you want it fast? A hosted tool, including our free preview, is fine.
For practical habits whichever route you pick, read how to keep meeting recordings private and how to transcribe sensitive meetings.