Local AI for your work: the model runs on your own computer

Local AI means: the AI model sits on your own computer and runs there. There is no service in the background receiving your input. You speak or select a text, the model processes it on your processor or graphics card, and the result appears in your program.

Technically, such a model is a large file full of numbers. A program loads it into memory and computes with it. None of this needs an outside connection: once installed, local AI works without internet, on a train or in an isolated network too.

The difference from cloud AI is not in how you use it, but in where your data goes. What that means legally is covered on our page about GDPR-compliant AI.

A model can only be run locally if its weights are openly available. Such open models are published as files and then run on your own hardware. To fit into the memory of an office PC, they are made smaller: the numbers in the model are stored more coarsely, the file shrinks, and the quality stays good enough for writing work.

With SpeechToWork you do not have to choose a model or put anything together. The software brings two parts: speech recognition that detects 25 European languages automatically, and a language model that puts the recognised text into shape. On start it checks your hardware and picks the right size itself.

You need no server and no special hardware. SpeechToWork runs on a current PC or laptop with Windows 10 or 11; 16 GB of memory is recommended. Installation takes a double click, and you need no administrator rights, neither for that nor for daily work.

• 64-bit Windows 10 or Windows 11.
• One-off model download on first start: 4 to 6 GB, depending on the computer.
• With a dedicated graphics card with 6 GB or more, the finished text appears in your program about one second after you release the key.
• On computers without a graphics card it takes a few seconds. A graphics card is a speed advantage, not a requirement.

After that, only the licence check needs a connection. SpeechToWork is available for Windows; a version for macOS is planned.

A local language model is strong when all the information it needs is already in the text. In office work, that is usually the case:

• Tidying up dictations: filler words and self-corrections disappear, numbers and punctuation are right.
• Shortening selected text, making it more formal or translating it.
• Writing email replies from a few key points.
• Summarising long messages and pulling out tasks.
• Taking meeting minutes, with decisions and action items.
• Reading invoices and screen areas.

Which of these tasks pay off first in a business, and what a workstation costs, is covered on the page AI for business. Local AI for law firms is described on the law firms page.

Working with your own data does not mean the model is trained with local AI. You show it the data and it works with it: you drag a PDF in and ask questions about it, you select an email and have the tasks pulled out, you add client names and technical terms to the dictionary so they are spelt correctly.

The model itself stays unchanged. It is downloaded as a file during setup and not changed afterwards. Your content never goes into any training run, because it never leaves your computer.

A model that runs on an office PC is smaller than the largest models in data centres. It has less general knowledge and is not suited to research or long creative texts. It cannot know today's news, because it does not look anything up.

That is why SpeechToWork uses local AI where it is reliable: tidying up, rephrasing, summarising, extracting. In addition, a check compares the language model's output with the recognised text. If the result looks implausible, SpeechToWork inserts the recognised raw text. Your dictation is never lost.