Whisper vs Paid Transcription: What You Get
What the open Whisper model gives you, what paid transcription adds, and what running it yourself costs in setup and hardware. No accuracy claims.
The Scriptorium team ·
Whisper gives you the speech recognition; a paid service gives you everything around it. The open model turns audio into text and timestamps for free. Paid services add things the model does not include: an editor, shared workspaces, human review, guaranteed turnaround, integrations and, usually, speaker labels. Whether those are worth money depends on whether you need them. We make a Whisper-based product, so weigh our view with that in mind, but this comparison sticks to what the projects and vendors say on their own pages.
What Whisper is, and what it is not
OpenAI's Whisper README says the code and model weights are released under the MIT licence. It lists six model sizes, four of them with English-only versions. The listed sizes run from tiny (39M parameters, about 1 GB of VRAM) to large (1550M parameters, about 10 GB) and turbo (809M, about 6 GB). The README says the transcribe() method reads the entire file and processes it with a sliding 30-second window, and that the turbo model is not trained for translation.
Two things the README does not do: it gives no language count (it points to a tokenizer file for the list) and it does not describe speaker diarization, meaning working out who spoke when. Whisper is a model and a command-line tool, not a product with an editor or accounts.
faster-whisper, what we run
Our server uses faster-whisper, which describes itself as a reimplementation of OpenAI's Whisper model using CTranslate2. It is a separate project under the MIT licence. Its maintainers claim it is up to 4 times faster than openai/whisper for the same accuracy and uses less memory; that is their own benchmark claim, not something we measured. It runs on CPU or on an NVIDIA GPU with CUDA 12 libraries, lists word-level timestamps, and mentions diarization only for community projects built on top of it. The no-account preview on our site uses the small model.
What a paid service adds
Rev's pricing page is a useful example because it sells both kinds. It lists human transcription from $1.99 a minute with a turnaround of 12 hours or less, and AI transcription as plan minutes: a free tier of 45 minutes per user, and paid plans from $25.49 per seat a month on its annual option. Other vendors price differently; check theirs.
What that money typically buys, and what a bare Whisper install does not give you:
- Human review. A person checks and corrects the text. This is the one thing that can make a transcript fit for legal or publication use, and no AI model can replace it.
- An editor and collaboration. Several people can open, edit, comment on and share one transcript in a browser.
- Turnaround and support. A stated turnaround or service level and someone to ask when it goes wrong.
- Speaker labels. Many services label speakers as a standard step. With Whisper you add a separate diarization tool and check its output.
- Integrations. Hooks into video calls, storage and other tools, so recordings arrive without you uploading them.
What doing it yourself costs
Free software is not free of cost. Plan for these:
- Setup time. Installing Python, a model and, for speed, GPU drivers. A desktop app such as Buzz or MacWhisper removes most of this.
- Hardware. The README's VRAM figures are the starting point: about 10 GB for large. CPU-only runs work but are slower. Long files take real time on modest hardware.
- Maintenance. You update models and dependencies, and you are the one who debugs a failed run.
- Missing pieces. Editing, speaker labels, subtitle formatting and storage are yours to assemble.
What you get in return is privacy, because the audio never leaves your machine, and no per-minute fee.
Where Scriptorium sits
Scriptorium is the middle option: the same family of open models behind a hosted editor, so you skip the setup. The no-account preview takes one file at a time, up to 15 minutes and 25 MB, with 3 tries a day, and deletes the file after 24 hours. Exports are TXT, timestamped TXT, SRT, VTT, Markdown, JSON and CSV, in 33 languages plus automatic detection. What we do not have: human review, an API, integrations, or burned-in captions. If you need any of those, a paid service is the better choice. If you need the audio to stay on your own hardware, you can self-host Scriptorium, and our self-hosted vs cloud article lays out the trade-off.
How to decide
| If you need | Pick |
|---|---|
| The cheapest route for occasional files | A free app or a free preview; correct the text yourself |
| Audio never leaves your computer | A desktop Whisper app or self-hosting |
| Court, publication or contract-grade text | Human review from a paid service |
| A team editing the same transcript | A paid service with a shared editor |
| Recordings arriving automatically from calls | A service with integrations |
On accuracy: the same model can be good on one recording and poor on another, and we know of no honest single number. See how accurate is AI transcription and test on your own file.
Questions worth asking any paid vendor
- Who reviews the text, and is there a human-reviewed tier? What does it cost per minute?
- Where is my audio stored, for how long, and can I delete it myself?
- Is my audio used to train a model? Ask for the answer in writing.
- What does the export include: subtitle files, timestamps, speaker names?
- What happens to my transcripts if I stop paying?
- Is there a published turnaround, and what is the remedy when it is missed?
The answers matter more than the price. A cheap service that keeps audio indefinitely may cost more, for some recordings, than a dearer one that deletes it.
You can mix the two
Many people use the free route for drafts and a paid human review for the few recordings that must be right. Others transcribe locally for anything confidential and use a hosted tool for the rest. Nothing about Whisper stops you using a paid service when a job needs one, and nothing about a paid service stops you checking its output against a free draft.
A short example of the trade-off
Take a one-hour interview you will quote in an article. With a free local app you pay in time (setup, waiting for the run) and get a draft you correct yourself. With a hosted tool like ours you skip setup and correct the same draft in a browser. With a human-reviewed service you pay per minute and get a document someone has already checked, which you should still skim. Which is right depends on how costly a wrong quote is, not on which model sits underneath.
Checked on 2026-10-24. OpenAI Whisper README, faster-whisper README and Rev's pricing page, each read on its own page. Rev's accuracy claims are deliberately not repeated here. Pricing of other vendors was not checked. Menus, prices and limits change; confirm on the vendor's page before you rely on them.