modelscope/FunASR ? reverse-engineered prompt

I want to build a comprehensive Python toolkit for speech recognition. The main goal is to take audio and produce accurate text. It should support both offline transcription of audio files and real-time transcription from a live stream.

A key feature is handling conversations with multiple people, so it needs to be able to identify who is speaking and when (speaker diarization). It should also be smart enough to automatically add punctuation to the final text to make it readable.

Let's build this using PyTorch. It should be designed to easily use pre-trained models from places like Hugging Face or ModelScope, including models like Whisper or other high-performance ones. The toolkit should also include other useful audio features like Voice Activity Detection to know when someone is actually talking, and maybe even speaker verification. Make sure it supports a wide variety of languages, not just English.