Whisper

Whisper is a speech recognition model from OpenAI, released with its code and model weights under the MIT license. You install it with pip, add ffmpeg, and run it from the command line or from Python. It comes in six sizes, from tiny (39M parameters) to large (1,550M), plus turbo, a faster version of large-v3 that is the default. The multilingual models transcribe many languages and can translate speech into English.

It is free and runs locally, so audio never has to leave your computer, but you need to be comfortable in a terminal, and the large model wants about 10 GB of GPU memory. Turbo is not trained for translation, so use medium or large for that. Apps such as MacWhisper and Superwhisper put a friendlier interface on Whisper models.