Whisper locally: transcribe and analyze government meetings without cloud

You are here: 2/7 · Part of the base flow in In contact with authorities

Why Local Transcription

After meetings with authorities, details are easy to lose. With speech-to-text, you get searchable text that strengthens your position when comparing with the authority's notes. In sensitive cases (children, health, social services, migration, police), the audio should not be uploaded to unknown cloud services. Run Whisper locally so the file stays on your computer.

Whisper is OpenAI's model for speech to text (not speech synthesis). It can be run entirely offline with ready-made apps or command-line tools.

Workflow

1. Record the meeting where you are participating (see the recording guide).
2. Save the original file unchanged. Make a working copy for transcription.
3. Transcribe locally with Whisper (steps below).
4. Review and manually correct important quotes against the audio.
5. Summarize decisions, promises, deadlines, and unclear points – preferably with local AI analysis (below).
6. Securely save both audio and text. Compare with the authority's notes and request corrections if needed.

How to Run Whisper Locally

Choose the path that suits your computer. All three keep the file offline unless you activate cloud features yourself.

Option A: Mac – MacWhisper or Aiko (easiest)

1. Install MacWhisper (godtone.com) or Aiko from the App Store / manufacturer.
2. Download a model locally (start with small or medium for Swedish; larger models often provide better quality but require more RAM).
3. Drag in the working copy of the audio file. Select language Swedish if available, otherwise auto.
4. Export as text or SRT with timestamps. Save next to the original recording.
5. Check in the app's settings that cloud/sharing features are turned off.

Option B: Windows / Linux – Whisper.cpp or Faster-Whisper

Whisper.cpp (C++ port, fast and local):

1. Get the project from the official GitHub repo ggerganov/whisper.cpp and build according to the README for your OS.
2. Download a model, e.g., ggml-medium.bin or ggml-small.bin, to the model folder.
3. Convert audio to 16 kHz WAV if the tool requires it (many guides show ffmpeg).
4. Run approximately: ./main -m models/ggml-medium.bin -l sv -f mote.wav -otxt
5. Open the text file, correct names and authority terms, save as YYYY-MM-DD_authority_transcript.txt.

Faster-Whisper (Python, often easier on Windows with GPU):

1. Create a virtual environment and install the package faster-whisper according to the project's documentation.
2. Run a short script that loads the model medium or small with device="cpu" (or CUDA if you have a GPU) and compute_type that suits your machine.
3. Set language to sv. Print segments with start time so you can quote “at 00:14:20 the officer said …”.

Option C: Buzz, Whisper Desktop, and Similar

Apps like Buzz (open source) can run Whisper locally on multiple platforms. Always choose “local model” / offline – not cloud transcription – when the case is sensitive.

Local Analysis of the Text (Without Leaving the Computer)

Once you have a transcription, you can structure it locally. Avoid pasting the entire text into ChatGPT or other cloud services if it contains children's information, health, protected personal data, or criminal suspicions.

• Completely offline LLM: Ollama, LM Studio, or GPT4All – steps, performance requirements, and prompts in Local LLM: analysis without third parties.
• Computer is not enough: temporary isolated GPU (e.g., RunPod Secure Cloud) can be the next best option – the same guide explains when it is reasonable and how to delete afterwards.
• Without AI: manually mark in the text: decisions, promises, deadlines, incorrect statements, what you should request in writing.

Short example prompt (more in the LLM guide):

You are a legal language assistant. Summarize this authority meeting in simple Swedish. Divide into: 1) decisions or notices, 2) promises from the authority, 3) things I need to do, 4) deadlines, 5) unclear points I should request in writing. Do not change quotes. Mark uncertain parts.

Minimal List: What You Always Want to Get Out

• Decisions and what is to happen next
• Who is responsible (officer, manager, school/social services)
• Which documents the authority bases its decisions on
• Deadlines (supplementation, reconsideration, appeal)

Legal Value

AI transcription is not an official document but can serve as supporting evidence along with the recording. Always save the original audio. The transcription is a work aid – quotes to be used in corrections or reports should match the audio.

Privacy

• Run a local model; turn off sync to iCloud/OneDrive for the folder with sensitive recordings if possible.
• Do not share the entire audio file if a short quote or a written summary suffices.
• Delete working copies from any cloud services if you have used them at any point.

Common Mistakes

• You only save the text but not the original file (difficult to substantiate in case of a dispute).
• You do not make a summary and lose track of "what was decided."
• You never compare with the authority's notes and miss errors.
• You upload entire social services or healthcare audio to a free cloud AI.

FAQ

Is AI transcription legally binding?
No, but it can support your evidence along with the audio file.

Which Whisper model should I choose?
Start with small or medium for Swedish. Upgrade if the quality is too poor and the computer can handle it.

Can AI help analyze the content?
Yes – locally with Ollama/LM Studio, or via an isolated GPU if the computer is insufficient. See the guide for local LLM and RunPod. Use the result for follow-up emails and correction requests, not as legal advice.

For Those Reading Under Stress

The guide is intended as support when you already have a lot on your mind. You do not need to read everything at once: start with the section that matches where you are today and return when the next step becomes relevant.

Cases with authorities rarely improve if you handle everything alone. Ask for help from a support person, lawyer, or relative if you can – especially before decisions with long-term effects.

Common Mistakes That Make the Case More Difficult

• Relying only on verbal promises without short written confirmation or follow-up emails.
• Waiting to request documents until the conflict is already locked – then it becomes harder to show what was said.
• Writing so long and emotionally that the authority stops responding to the substantive issue.
• Missing deadlines for reconsideration or appeal because the decision was not read verbatim.

Mini-Checklist Before Leaving the Guide

1. Do I have the date, diary/case number, and name of the person handling the case (if known)?
2. Do I know what the next step is: respond, request, remind, appeal, or report?
3. Have I saved emails and attachments in a place where I can find them in a week?

Basic Guide: In contact with authorities