Running it inside your own network

Why a free tool still isn't always usable

Not because of anything it does. Because of four things it can't be:

• The domain isn't yours. Nothing is uploaded — your browser enforces that, and you can check it in thirty seconds. But a rule that says confidential material is handled only on systems the organisation controls is a rule about whose domain it is, not about where the bytes went, and it isn't wrong to have one.
• There's nothing to file. No contract, no invoice, no named supplier, no signed description of behaviour. A regulated team often cannot adopt a tool that produces no paperwork, however well it works.
• It changes without warning. Every time we deploy, the tool you assessed is replaced by a different one. A version you validated needs to still be there next year.
• Nobody is obliged to answer. Goodwill is not a support commitment, and it is worth nothing in the middle of a week you can't move.

What a licensed copy is

• A folder of static files you host yourself — the page, the runtime and the model — served by your own web server, on your own network.
• The model files travel inside the bundle instead of being fetched on first use. That is the part that makes a machine with no route to the internet work at all.
• A version that doesn't move until you ask it to.
• A licence agreement, an invoice, and a written statement of what the software does — the things a compliance file can actually hold.
• A named person to email, and a written answer to your security questionnaire.

What your web server has to do

Little, but one part is easy to miss and expensive to miss, so it's worth stating before anyone quotes you a day of work for it:

• Serve static files over HTTPS. There is no application server, no database, and no process of ours running on your infrastructure.
• Send two headers — Cross-Origin-Opener-Policy: same-origin and Cross-Origin-Embedder-Policy: require-corp. Without them the browser withholds the shared memory the model needs to use more than one processor core, and transcription runs several times slower for no visible reason.
• Allow a browser tab about 2 GB of memory, or about 3 GB when it uses the graphics card. Both figures are measured.
• No outbound access. Not for the model, not for licensing, not for telemetry. There is no call home to allow, because there is nothing to call.

What it doesn't change

The same code and the same model, so the error rate on the measurements page is the error rate you get. Nothing is held back from the free tool to make this version worth buying — the free one isn't crippled, it's the same thing in a place your policy may not permit.

Nothing is added either. No licence check, no seat counting, no usage reporting. A tool that argues your recording shouldn't leave your machine does not get to ship a version that reports back, and a paid copy that phoned home would be a worse product than the free one.

What we won't claim

• This is not a compliance certification. We'll describe the software accurately and answer a questionnaire in writing. Whether that meets your obligations is your advisers' call, and any supplier who tells you a piece of software settles it for you is overselling.
• Speaker labels aren't shipped — in either version. If "who said what" is the requirement rather than a preference, this isn't ready for you yet.
• It is not as accurate as a person. On recordings where people talk over each other it misses words nobody could have caught from a single line of text. If you need verbatim, hire a human. The numbers are published either way.

Is it worth asking about?