whisper.cpp: transcribing audio without sending it anywhere

Meetings, calls, voice notes, interviews.

Almost every company produces hours of audio.

And almost none of them do anything with it.

What it is

whisper.cpp is a C/C++ implementation of Whisper, the speech recognition model published by OpenAI. It transcribes audio to text on your own computer, offline, without sending the recording to any external service.

It runs on a huge range of platforms, from Apple computers to Linux and even small devices. And it comes from the same family of projects as llama.cpp.

Why I find it interesting

Because it turns something that gets lost into something you can use.

What gets said on a call with a client is gold: what worries them, what they ask for, the exact words they use to describe their problem. But if it isn't written down, it's forgotten in two days.

Transcription was always possible. It was just expensive, slow, or meant sending private conversations to a third-party service. With a local tool, the first two barriers drop and the third disappears.

Information that isn't written down doesn't exist. All that exists is the memory that someone said something.

Ideas for a small company

• Sales calls: transcribe them (with the other person's permission) and review later which questions keep coming up.
• Voice notes: talk instead of typing and have the text ready to organise.
• Internal videos and training: generate subtitles or a written summary.
• Customer interviews: keep their exact words, not your interpretation of them.

And if you combine it with a local language model, the next step is obvious: summarise, classify or pull the to-dos out of each conversation. All without the audio leaving your machine.

What to keep in mind

It isn't perfect. With a lot of background noise, several people talking at once or strong accents, it makes mistakes. Always check what matters.

And there's a part that isn't technical: if you record someone, tell them. Running the tool locally doesn't change what the law says about recording conversations.

What this repository teaches

That often the best idea isn't to build a new model, but to make an existing one work everywhere. Whisper was already published. whisper.cpp made it fast, light and portable.

And that took it from the lab to any computer.

P.S. Try transcribing your next call with a client and read it the next day. You'll be surprised at what you don't remember hearing.