Should you just run whisper.cpp yourself?

Most comparison pages on the internet exist to explain why the alternative is worse. This one cannot, because the alternative is the thing our app is built on.

whisper.cpp is Georgi Gerganov's C/C++ implementation of OpenAI's Whisper. It is MIT licensed, free, actively developed, and very good. Offline Transcription uses it as its engine. When it gained built-in voice activity detection, our app gained it too, because it is the same code.

So the honest question is not "which is better". It is "do you want to do the assembly yourself".

Run it yourself if… #

You are comfortable in a terminal. The whole thing is ./build/bin/whisper-cli -m model.bin -f audio.wav. If that sentence is unremarkable to you, you are ten minutes from a transcript and you should not pay anyone for this.

You want to script it. Watch folders, cron jobs, piping transcripts into other tools, transcribing on a server — this is where a CLI is not just cheaper but genuinely better. No GUI competes with a shell loop for automation.

You want every knob. Beam size, temperature fallback, custom prompts, alternative models, entirely different quantisations. We expose the settings we think most people need; whisper.cpp exposes all of them.

Your budget is zero. Completely legitimate. It is free software and it works.

Use the app if… #

You do not want to build anything. Getting the fast path on Apple Silicon means compiling with the right flags and converting the encoder to Core ML with Python tooling. It is documented and it is not hard, and it is still an afternoon the first time and a re-run every time you update. We do that once and ship the result.

You want it on your phone. There is an example iOS project in the repository, and building, signing and installing it on your own device is a real Xcode task that expires with your provisioning profile unless you have a developer account. A universal app from the App Store is a different proposition.

The work is in the reviewing, not the transcribing. This is the honest core of it. Getting text out of audio is one command. Finding the passage you need, checking a name against the audio, fixing it, and exporting subtitles is where the actual time goes — and a CLI gives you a text file and wishes you luck. A segment browser where you click a line and hear that line is not something a terminal can do.

Other people need to use it. Colleagues, a research team, an editor. "Install Xcode command line tools and run make" is not a workflow you can hand to someone else.

What you are actually paying for #

Not the model — that is free and public. Not the engine — that is MIT licensed and public.

You are paying for the packaging: a Core ML encoder compiled and shipped, a quantised model bundled at a size that fits a phone, a queue, a language picker per file, a segment browser with audio playback, seven export formats, and long-file handling that keeps memory flat so a two-hour recording does not get your process killed on an iPhone.

That is a real amount of work, and it is also completely reasonable to decide you would rather do it yourself for free.

One thing worth saying about open source #

Because our app depends on it: if whisper.cpp is useful to you, whether through our app or directly, it is worth starring the repository and reading the issues. The reason local transcription got good on consumer hardware in the last couple of years is largely that project and the ggml library underneath it.

We track it closely — our write-up of how on-device transcription works is essentially a description of what whisper.cpp does, because that is what is running.