The stream example in whisper.cpp solves this problem effectively by running the inference on a sequence of chunks and then keeping around the last chunk as con

The stream example in whisper.cpp solves this problem effectively by running the inference on a sequence of chunks and then keeping around the last chunk as context for the next sequence of chunks for the next inference. Ideally as many chunks as possible should be combined into a sequence as this would minimize the amount of redundant computation and edge effects should the speaker have an utterance that is more than 30 seconds long. But just doing 2 chunks at a time would also work though it would double the processing required.