Answering my own question, for the sake of posterity: Just found banana.dev, which allows you to essentially 1-click deploy Whisper. Then you can use their Py

Answering my own question, for the sake of posterity: Just found banana.dev, which allows you to essentially 1-click deploy Whisper. Then you can use their Python library to connect and run inference. They use the GPU variant by default (not whisper.cpp), and you are being billed for GPU hours. By default, the model "hibernates" after 10 seconds of inactivity. Which, for me, results in inference times of around 15 seconds (most of which, I presume, is for "waking up" the model).