whisper

Whisper turns speech into text, and it set the bar the moment OpenAI released it. Feed it an audio file in almost any of 99 languages and you get back a transcript, optionally translated to English. The model weights and the code are MIT licensed, so you can run the whole thing on your own machine for nothing.

Running it yourself is a pip install and an ffmpeg dependency away, but the catch is hardware. The tiny model fits in about 1GB of VRAM and is fast and rough; the large model wants roughly 10GB and a real GPU to run at a sane speed. On a CPU it works, but you will wait. A newer turbo model is much faster for plain transcription, though it drops the translation trick.

For a one-off transcript, OpenAI's hosted Whisper API runs about half a cent per minute and saves you the setup. Run it locally when the audio is sensitive, when you are processing a lot of it, or when you just do not want a per-minute bill. Solo and small teams: local on a decent GPU is plenty. Higher volume: budget a GPU box and self-host.

The catch is that Whisper is a model, not an app. It does straight transcription, not speaker labels or live captioning out of the box. If you want a GUI with those niceties, look at buzz or vibe, which both wrap this exact model.