How to Transcribe Audio Offline on a Laptop

What runs where

Speech-to-text has two very different forms on a laptop. Built-in live captions listen to whatever is playing and show text on screen as it happens. Microsoft's support guide says Windows 11 live captions process audio entirely on the device and work without an internet connection, and Google's Chrome help says its Live Caption audio and captions never leave the device. These are designed for reading along, not for producing a tidy saved document. A transcription app takes a file, runs a speech model over it and gives you text with timestamps you can edit and export. That is the tool for interviews, lectures and meetings you need to quote from.

Hardware: less than you might think

Speech models are much smaller than writing models. The research paper that introduced one widely used family of open speech models, published on arXiv in 2022, describes models from tens of millions up to about 1.5 billion parameters, trained on a very large and varied collection of audio so they cope with accents and background noise. The small and medium sizes run on an ordinary laptop with 8 GB of memory; the largest are happier with 16 GB and a graphics chip. A neural processing unit can help with live captioning, as the guide to what an NPU does for tasks like captions explains, but a file transcription app mostly uses the processor and graphics.

Transcribe a recording step by step

1. 01

   Convert or locate the recording

   Most apps accept common audio and video formats directly. If yours does not, export the recording as a WAV or MP3 file from the recorder or phone app.

2. 02

   Install a transcription app from its official source

   Choose an open-source desktop app that says it runs models locally. Download it from the project's own website or code repository, not a download portal.

3. 03

   Pick a model that fits your memory

   Start with a small or medium model on 8 GB of memory and move up only if accuracy is not good enough. Set the recording's language rather than relying on auto-detection.

4. 04

   Test with the network off

   Switch off Wi-Fi and transcribe a one-minute clip. If it works, the model is running on the laptop. If the app complains, look for any cloud, account or sync option in its settings and turn it off.

5. 05

   Run the full file on mains power

   Transcription keeps the processor busy for the whole job. Plug in, keep the vents clear and leave the laptop to work.

6. 06

   Check the transcript against the audio

   Use the next section. Then export as plain text or subtitles, and keep the transcript beside the original recording.

Check accuracy before you quote anything

Researchers score speech recognition by word error rate: the substitutions, deletions and insertions needed to turn the machine's text into a correct transcript, divided by the number of words. Even a low average hides the errors that matter most to a writer. Check these every time:

• Names, places and numbers. Models guess unfamiliar names phonetically and often mishear figures. Search the transcript for every number and check each one against the audio.
• Direct quotes. Before quoting anyone, listen to the exact passage again. A transcript is a finding aid, not the record.
• Silences and music. Speech models can produce text that was never said, especially over long pauses, music or crosstalk. Skim any passage where the timestamps jump.
• Speaker changes. Many local apps do not separate speakers reliably, so mark them yourself.

Privacy and storage

Recordings of interviews are often the most sensitive files on a journalist's or student's laptop. Keep them in one folder, encrypt the drive with the operating system's built-in disk encryption, and keep recordings in more than one place so a dead drive does not take the only copy. If the laptop also has built-in AI assistants you do not use, you can switch off the AI features built into your laptop. The same local approach works for drafting: you can set up a private writing assistant on the same laptop. For other practical walkthroughs, browse more laptop how-to guides.

Common questions

Is offline transcription as accurate as a cloud service?

On clear recordings the best local models come close. Cloud services may do better on very noisy audio or with speaker labelling, at the cost of uploading the recording.

Can Windows live captions save a transcript?

They are designed to be read on screen as audio plays. For a saved, editable transcript with timestamps, use a transcription app.

Does it work for languages other than English?

Many open speech models are multilingual, with accuracy that varies a lot by language. Set the language manually and check a sample first.

Is the transcript good enough to quote from?

Only after you have checked the passage against the audio. Treat the transcript as a map to the recording.