I had a kickoff meeting for a new project today. Instead of taking notes, I recorded the conversation, with everyone's permission, and transcribed it afterwards. The difference in the quality of the conversation was immediately obvious.
The original version was a tweet in Hebrew.
Why taking notes costs you the meeting
When I take notes, I am splitting my attention. Half my brain is busy summarizing the previous sentence instead of listening to the current one. That is not a small tax in a first conversation with a client, which is exactly where the useful, unrepeatable context lives.
This time I was fully in the conversation. And afterwards I had everything that was said, not just the parts I managed to catch.
Notes give you what you managed to write down. A transcript gives you what was actually said.
What I run: mlx-whisper with the ivrit.ai Hebrew model
My meetings are in Hebrew, and stock Whisper on Hebrew is mediocre. I use mlx-whisper with the Hebrew model from ivrit.ai, which is a finetune of Whisper large-v3 on Hebrew. It is a completely different quality level from the general model.
The 44 minute conversation transcribed in a few minutes.
The script
I wrote a small script that asks for a path to an audio file and returns a text file with timestamps and paragraph breaks. Nothing clever, it just removes the friction so I actually use it after every meeting instead of meaning to.
No speaker diarization right now. For what I need it is not critical, since I know who was in the room and the content is what matters. I may add it later.
The code is on GitHub: razamit/audio_transcriber.
What you need to run it
- A Mac with Apple Silicon (mlx-whisper runs on MLX).
- 16GB of RAM is the comfortable minimum. 8GB will work, but it will be tight.
Why local matters here
This one is not about cost. A client kickoff is the single most sensitive recording I make in a week: names, numbers, internal politics, things people say out loud precisely because the room is small. Running the transcription on my own machine means that audio never becomes somebody else's training data or somebody else's breach.
It is the same test I apply everywhere: use a local model where the data is the reason, not where the benchmark is the reason. I wrote about the rest of that line in where local LLMs actually earn their keep.
If you are working out which parts of your workflow should stay on your own hardware and which should not, that is a conversation I have with clients constantly, and it is part of AI consulting.
FAQ
What is the best local model for transcribing Hebrew audio on a Mac?
I use mlx-whisper with the Hebrew model from ivrit.ai, which is a finetune of Whisper large-v3 on Hebrew. The quality gap against stock Whisper on Hebrew is not subtle. mlx-whisper runs on Apple Silicon through MLX, so everything stays on the machine.
How long does it take to transcribe a meeting locally?
A 44 minute conversation transcribed in a few minutes on my Mac. You need Apple Silicon, and 16GB of RAM is the comfortable minimum for the large model. 8GB will work, but it will be tight.