Prepare the official software/model for your platform, run one permitted short recording, and inspect the transcript against the audio. Its output is source material for minutes, not confirmed team decisions.
In this guide
When does a local transcription setup fit the task?
Consider this route when recordings recur and someone can maintain the executable, model files, and output folders. If your meeting system already provides a usable transcript, export it and spend the effort on corrections and the handoff.
For a first trial, choose a permitted excerpt containing the names, numbers, and speaking style you need to handle. Record setup time, processing time, correction effort, and unresolved passages. Use those observations to decide whether the local route fits your work.
You want to keep transcription on your own computer
Use the local executable and models described below. If you later paste the transcript into a hosted AI service, check that separate service's handling and permission requirements.
You need dependable named action owners
Check speaker identity and commitments from the recording or participants. Treat diarization labels as draft aids requiring review; the basic commands below request text/SRT output rather than a named action register.
Prepare the executable, model, and a short audio file
Follow the official repository's release/build and model instructions for your operating system. A source build requires the documented build tools; Windows uses a compatible C/C++ toolchain. Keep the software version and model file name in your run notes.
This local path uses your computer and model files. It does not require sending audio through a site-funded API. Model downloads, storage, memory, and processing time still belong to your setup.
Use a predictable audio format for the first run
For the first run, prepare a mono 16 kHz 16-bit PCM WAV using the command below. Check your installed CLI help before choosing another input format. Run FFmpeg in your own prepared environment with a permitted input file.
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le meeting-16k.wav
Windows: run the prepared executable
Replace every example path below with your actual executable, downloaded model, input WAV, and output stem. Use a new output stem when keeping a previous run.
& "C:\your-folder\whisper-cli.exe" -m "C:\your-folder\ggml-base.en.bin" -f "C:\your-folder\meeting-16k.wav" -osrt -otxt -of "C:\your-folder\meeting"
macOS/Linux: run from the prepared project folder
After following the official build/model instructions, use the executable location for your build. This example uses the documented project layout.
./build/bin/whisper-cli -m models/ggml-base.en.bin -f meeting-16k.wav -osrt -otxt -of meeting
Inspect outputs and resolve uncertain passages
The requested outputs are text and SRT files under the chosen output stem. Open the SRT beside the recording to check positions, names, numbers, and negations. A timestamp does not by itself identify a speaker.
For a model-path error, check the actual model file. For an input error, check the WAV preparation and audible content. For an inaccurate passage, listen and retain uncertainty instead of asking the summary to guess.
- Keep the original recording and raw transcript.
- Record important corrections with their audio positions.
- Use the corrected transcript to prepare minutes, then ask the team to confirm actions and dates.
References
Use it in a task
AI Meeting Minutes for Project Leads: Transcribe, Check, Summarize
Continue the task
Use the corrected transcript as evidence for the next task, retaining uncertain names and conditional commitments.