How to transcribe audio without uploading it to anyone's server

Every free transcription site works the same way: you hand over the recording, it comes back as text. For a podcast episode that is fine. For a research interview, a client call, a patient conversation or anything with a confidentiality clause on it, “hand over the recording” is the part you are not allowed to do.

The useful thing that changed in the last couple of years is that you no longer have to. A speech model good enough for real transcription runs on an ordinary laptop, offline, for free.

what uploading actually commits you to

A recording of someone talking is personal data, and often special-category data - health, opinions, employment, a name said out loud. Sending it to a transcription service means:

• A processor you have to name. If you work under GDPR, that service is a data processor and belongs in your records, usually with a data processing agreement behind it.
• Retention you do not control. Most services keep the audio and the transcript until you delete them, and some keep backups longer.
• A promise you passed on. If you told an interviewee the recording stays with you, a free web converter is the moment that stopped being true - whatever the service does with it afterwards.

None of that is an accusation against any particular service. It is the ordinary consequence of the file leaving your computer, and the reason ethics boards, DPOs and in-house counsel ask about it.

what runs locally, and how well

The model to know about is Whisper, released openly by OpenAI and runnable through whisper.cpp on a normal CPU. It handles a long list of languages, produces word-level timestamps, and can translate into English. This is the same family of model the paid services run - the difference is where it runs, not what it is.

Two things to be realistic about before you commit a deadline to it:

Speed

On a four-core laptop CPU, the large turbo model transcribes roughly in real time - we measured 33 to 37 seconds for 24 seconds of speech. So a one-hour interview is about an hour of your laptop working, which is fine overnight or over lunch and painful if you need it in five minutes. Smaller models are several times faster and noticeably worse on names. A machine with a decent GPU changes this picture completely.

Speakers

Whisper transcribes speech; it does not tell you who is talking. Cloud services often add speaker labels on top. If your workflow depends on “Interviewer:” and “Participant:” being filled in for you, a local transcript will need that pass by hand - for a two-person interview that is usually minutes, but it is honest work you should know about in advance.

getting a usable transcript, not a wall of text

Raw model output is one long block, which is unreadable and unquotable. Three things make it work as a document:

1. Sentence or paragraph grouping. The model emits short cues that often end mid-sentence, so they need regrouping into sentences and paragraphs before anyone can read it.
2. Timestamps you can cite. For a quote you will want to go back to, a timestamp every paragraph beats a timestamp every four words.
3. A vocabulary. Names, places, product names and jargon are where every recognizer fails. Whisper accepts an initial prompt - give it the participant’s name and the six terms your field uses and the transcript stops being littered with near-misses.

If you need subtitles rather than a document, the same run gives you an .srt or .vtt file, because the timings are already there.

the three ways to do it

The command line is genuinely fine if you transcribe something twice a year. The reason most people end up wanting an app is not the transcription - it is the formats, the regrouping, the vocabulary and not re-reading a manual every time.

how to check the claim

Any tool that says “runs locally” can be tested in ten seconds: turn off wi-fi after the model has downloaded, and transcribe something. If it works offline, the audio is not going anywhere. For a stronger guarantee, add an outbound block rule for the program in Windows Defender Firewall and try again. We wrote the same test up for dictation in dictating on Windows without the cloud.

Who wrote this. We build owntools, which includes the transcription tool described at the end. The speed figures above are our own measurements on one four-core laptop, and the speaker-labelling limitation is a real one we would rather you hear from us than discover at midnight.