Free transcription and dictation on your own computer: Whisper locally, no internet

Transcription services charge by the minute of audio. Dictation services charge a monthly subscription. And the two have something in common that almost never gets discussed: to work, your voice has to leave your computer and reach someone else's server.

There is another way. Whisper, OpenAI's transcription model, was released under an MIT licence and can be run on your own machine. This article explains what you actually need, what you gain, and the cases where the cloud is still better.

What "local" means

It means the model is downloaded once and runs on your processor or your graphics card. The audio is not uploaded anywhere. There is no API call, no per-minute cost and no need for a connection.

In practice it covers two different things:

• Transcribing: turning a recorded meeting, an interview or a video's audio into text.
• Dictating: speaking and having the text appear in whatever application you are using, replacing the keyboard.

What computer you need, without the marketing

This is the question that gets answered worst, so here are approximate and honest numbers.

Whisper comes in several sizes, and the size decides everything:

• tiny and base (around 75–150 MB): they run on almost any laptop, CPU only. They are fast and they get a fair amount wrong, especially with proper nouns and strong accents. Good for finding something inside an audio file, not for publishing.
• small (around 500 MB): the reasonable balance. With 8 GB of RAM it works on CPU, without much room to spare.
• medium (1.5 GB): the quality is noticeably better. On CPU it is slow; with a dedicated graphics card, comfortable.
• large-v3 and its turbo variants (around 3 GB): reference quality. Here you do want a dedicated card with 6 GB of VRAM or more.

As a practical rule: without a dedicated graphics card, stay on small. With one, use large in its turbo variant. On a Mac with an Apple chip things improve considerably, because the model can use the Neural Engine and runs noticeably faster than on pure CPU.

On speed: with a decent graphics card, an hour of audio transcribes in a few minutes. On CPU alone with the small model, expect something close to the length of the audio itself. It is not instant, but you launch it and leave it working.

The advantages

Local processing. Whisper can transcribe audio on the computer itself, without sending it to a cloud transcription service.

Zero cost per use. Services charge by the minute. If you transcribe four hours a week, that is a permanent monthly bill. Locally, the bill is your electricity.

No limits or quotas. You can transcribe an entire historical archive in one go. On a paid service, that same job is a budget decision.

It works offline. On a plane, on a shoot with no signal, or on hotel Wi-Fi, it keeps working.

It does not expire. The model is downloaded. If prices change tomorrow or the service shuts down, your installation stays exactly as it is.

The downsides, in detail

Quality depends on the model, and the model depends on your machine. This is the main one. A tiny on an old laptop makes mistakes a paid service does not. Comparing "local" with "cloud" without saying which model is being used is cheating.

The first time, there is a download. Between 500 MB and 3 GB depending on the model. Once only, but you have to do it.

It uses up your machine while it works. While transcribing, the computer runs slower and you hear the fan. On a laptop, the battery drops fast.

Diarisation is weaker. Separating who said what — "this was Ana, this was Luis" — is where commercial services have the biggest edge. It exists locally, but it is worse.

Bad audio is still bad audio. Neither the cloud nor a local model rescues a recording with echo, three people talking at once, or the microphone two metres away. Input quality matters more than the model.

Languages and accents are uneven. Whisper does very well in Spanish and English. In languages with less training data, or with very strong accents and a lot of technical jargon, the error rate goes up.

When the cloud still wins

With the same honesty:

• If your computer is modest and you need high quality, a paid service will give you a better result than a local tiny.
• If you need fine-grained diarisation with many speakers.
• If you transcribe very little, three audio files a year, it is not even worth downloading the model.
• If you need immediate results at high volume and your machine cannot keep up.

It is a decision per case, not a matter of picking sides.

What makes a transcript useful

Transcribing for the sake of it does little: you end up with a nine-thousand-word .txt you are not going to read either.

Where this changes is when the transcript goes into an index. If what was said in each video is searchable text, you can ask for "the interview where he talks about the process by hand" and land on the exact minute, months later, without remembering the date or the file name.

In Polimake Desktop transcription is free and local, precisely because your machine is the one running it. What you pay for — if you want it — is having that text and that material live in a shared archive, searchable by the whole team and reachable from your assistant. That is where there is storage behind it.

But the part where you transcribe on your own computer, without the audio leaving it, is free and will stay that way. Not out of generosity: because it costs us nothing.