Skip to main content

KNIFE – Knowledge In Friendly Examples Series: Systemic Thinking in IT & Digital Fabrication Level: Beginner Tags: claude-code transcription whisper ollama local-ai offline privacy audio mac

Local audio transcription and AI on top

Fastest: if you have VS Code and the Claude Code extension, tell Claude where the recordings are. It does the rest for you. This is way C below and it is how the transcripts in this KNIFE were made.

🎯 What it solves (purpose, goal)​

A colleague sends a ten-minute WhatsApp voice message and you need it as text. People with recorded lectures, workshops or meetings have the same problem. Online transcription services upload the recording to someone else's server. That is a poor fit for confidential conversations and personal recordings.

This KNIFE shows a workflow where neither the recording nor the text leaves your computer. The output is text plus timestamps relative to the start of the recording, so you can jump from the text back to the exact place in the original.

Two solutions and the easiest way​

By purpose there are two solutions (A and B). The second builds on the first. The third way (C) is not another technology but a way to run both A and B without IT skills.

Three ways: C tell Claude, A transcript, B transcript plus local AI

C Tell ClaudeA TranscriptB Transcript plus local AI
You wantto say only where the files aretext and its place in the originalto grasp the content fast
Outputas in A or B.txt, .srt, .jsonsummary, chapters, actions with time
ToolsVS Code and Claude Codemlx-whispermlx-whisper and Ollama
Effortone-time Claude Code setupa few minutes, about 1.5 GBa few GB more, slower
Statusverified, this is how it was madeverifiedAI step is a design, unverified
Fitscolleagues without IT skillsan occasional voice messagemany or long recordings

Details: solution A is under "How it solves it" and "How to use it", solution B under "Solution B" below.

⭐ Easiest: just tell Claude (no IT skills)​

This way is for people who do not want to install anything or use a terminal. You only need Visual Studio Code with the Claude Code extension. Claude does the rest for you: it installs the tools, transcribes the recordings and saves the outputs.

  1. Put the recordings in one folder on your computer.
  2. Open that folder in VS Code and start Claude Code.
  3. Tell Claude what you want. For example:

The voice notes from WhatsApp are in this folder. Transcribe them locally into text with timestamps, in English. Save the outputs in a subfolder called prepis. Do not upload anything to the internet.

  1. On first run Claude asks for permission to run commands. Allow them. Then you find the finished files (.txt, .srt and more) in the prepis subfolder.

This is how the transcripts behind this KNIFE were made. All it took was telling Claude where the files were, and finished outputs came back. Two more voice messages that arrived later were transcribed the same way.

One-time setup (the only complication)​

You do this once. After that you only write to Claude.

  1. Install Visual Studio Code (free).
  2. In VS Code open the Extensions panel (the squares icon on the left bar), search for Claude Code and install the extension from Anthropic.
  3. Sign in with your Claude account. Claude Code needs an account with access, for example a Claude subscription or an Anthropic Console account, under Anthropic's current terms.
  4. Open the folder with the recordings (File, Open Folder) and open the Claude Code panel.

What it needs. The first run needs internet to download the tools and the model (about 1.5 GB). The transcription itself is then done by Whisper on your computer and the audio file is not uploaded anywhere. It is verified on an Apple Silicon Mac. On another computer Claude will pick a different tool, but I did not try that.

⚠️ What Claude sees. Claude is a cloud model. It sees your prompts, command output and any file it reads. It does not read audio, but if you let it read the transcript, the text reaches it. If the content is confidential, say: "Do not read the transcript, just tell me where it is saved."

If you also want a summary, chapters and action items (solution B), tell Claude that too. That step is still a design.

🧩 How it solves it (principle)​

Solution A – transcript to text (verified)​

The recording passes through three steps, all local:

  1. Decoding. ffmpeg converts .opus, .m4a or video to mono audio.
  2. Speech recognition. The Whisper model turns the audio into text. It needs no internet once the model is downloaded.
  3. Segments. Every sentence gets a start and an end in seconds from zero of the recording.

Local audio transcription: flow from recording to KNIFE item

Different kinds of recording (voice message, lecture, workshop, meeting, dictaphone) are only different inputs to the same flow.

🧪 How to use it (application)​

The steps are verified on an Apple Silicon (arm64) Mac. You do not need Homebrew.

Fastest: one script​

Download the script tools/prepis.sh (it sits in this KNIFE's folder) and run it. On first run it installs everything it needs.

curl -O https://raw.githubusercontent.com/KNIFE-Framework/knifes_overview/main/content/docs/en/knifes/K000122-Local-Audio-Transcription/tools/prepis.sh
chmod +x prepis.sh
./prepis.sh en "lecture.m4a" # English transcript
./prepis.sh sk "voice.opus" # Slovak transcript

The result lands in a prepis/ subfolder next to the recording. Optionally add GLOSSARY="names, terms" to help Whisper recognise names. To see what the script does, or to run it in parts, follow steps 1 to 4 below.

ℹ️ The curl link starts working once this KNIFE is published on GitHub. Until then copy prepis.sh from the local repository.

1. Get the file​

A WhatsApp voice message is an .opus file. In WhatsApp Desktop right-click and choose Save as. From a phone, send it to yourself by AirDrop or e-mail.

2. Install the tools in their own environment​

python3 -m venv ~/whisper-env
~/whisper-env/bin/pip install mlx-whisper imageio-ffmpeg

mlx-whisper is Whisper optimised for Apple Silicon. The imageio-ffmpeg package brings ffmpeg without Homebrew. mlx_whisper looks for a program named ffmpeg on PATH, so link it:

ln -sf "$(~/whisper-env/bin/python -c 'import imageio_ffmpeg;print(imageio_ffmpeg.get_ffmpeg_exe())')" ~/whisper-env/bin/ffmpeg

The environment lives outside your projects, so it does not touch other Python installations.

3. Transcribe with timestamps​

export PATH=~/whisper-env/bin:$PATH
mlx_whisper "voice.opus" \
--model mlx-community/whisper-large-v3-turbo \
--language en \
--output-format all \
--output-dir transcripts \
--output-name voice_01

The first run downloads the model (about 1.5 GB). Later runs are offline. --output-format all writes every format at once:

FileWhat it containsWhat you use it for
.txtplain text without timesreading, copying into an e-mail or document
.srttext in segments with times, for example 00:01:04,420 --> 00:01:07,240find in which minute something was said, subtitles for video
.vttthe same as .srt, written for the websubtitles for web video
.tsvtable: start and end in milliseconds, textopen in Excel or LibreOffice, filter and sort
.jsonthe full Whisper output: text, segments, times and technical datafurther processing in code, for example by AI

Usually .txt is enough for reading, and .srt when you need times. All files are the same transcript in different notation and carry the same name as the recording with a different extension.

All times are from the start of the file. The mark 00:04:12 means 4 minutes 12 seconds after the recording began.

4. Check the result​

In practice accuracy was around 90 % on Slovak voice messages with weak audio. Errors are mostly diacritics, names and technical terms. Treat the transcript as a good draft to edit, not as final text.

💡 Tips and notes​

  • File names with dots in the time. WhatsApp files are named like WhatsApp Audio 2026-11-15 at 09.51.17.opus. When run over several files at once, the outputs overwrote each other into one file. Run the program one file at a time. The prepis.sh script does this itself and names the outputs like the input, with a different extension.
  • Language. Set --language to the spoken language, for example sk or cs. Without it Whisper guesses, and on short recordings it can guess wrong.
  • Accuracy or speed. whisper-large-v3-turbo is the fast compromise. mlx-community/whisper-large-v3-mlx is more accurate but slower.
  • No terminal. The MacWhisper app does the same by drag and drop and also runs locally.
  • Privacy. The internet is needed only once, to download the packages and the model. The transcription itself runs offline.
  • Other platforms. The mlx-whisper route works only on Apple Silicon Macs. On Windows or Linux use faster-whisper or whisper.cpp; the flow is the same.

🤖 Solution B – transcript plus local AI (design)​

Transcription is only the first step. The second step lets a local AI turn the text into a summary, chapters with time and a list of action items, with nothing leaving the computer.

Full context: VS Code, Claude Code and the local Whisper plus AI chain

Who does what. VS Code runs the Claude Code extension. The cloud Claude model helps write and run the script. The recording itself is then processed by local tools: Whisper for the transcript and a model in Ollama for the summary. Once the chain is set up, Claude is no longer needed.

What is verified and what is not. The transcription step (prepis.sh without a summary) is verified. The local AI summary (SUMMARY=1), folder watching and the name glossary are a design. The script is ready, but I did not test a local model. Slovak quality of small models in particular must be tried.

How to try it:

# 1. Install Ollama (https://ollama.com) and pull a model, a few GB
ollama pull qwen2.5:7b

# 2. Transcript and summary in one command
SUMMARY=1 GLOSSARY="Kazicka, KNIFE, BaZi" ./prepis.sh en "lecture.m4a"

The summary is saved as prepis/<name>.summary.md. The model receives the .srt, so chapters and action items carry a time from the start of the recording.

⚠️ Privacy. Claude Code uses a cloud model. It sees your prompts and everything Claude reads, including terminal output and files. If the recording is confidential, run the script yourself in a terminal and do not ask Claude to read the transcript. Then not even the text leaves your computer.

✅ Value / Summary​

  • A colleague without IT skills can do the transcription alone: VS Code, the Claude Code extension and one sentence are enough.
  • A recording becomes text with the exact position in the original, and the audio is not uploaded to the cloud.
  • One flow covers voice messages, lectures, workshops and meetings.
  • Timestamps let a KNIFE link to a specific minute of the recording.
  • The result needs a read-through, especially names and technical terms.
  • Optionally a local AI turns the recording into a summary, chapters and action items with time (design).

Sources​

⬅ KNIFES – Overview • List • Details