microsoft/VibeVoice ? reverse-engineered prompt

Build me an advanced voice AI application that can handle both transcription and speech generation.

For the transcription part, I need it to process long audio files, like an hour-long meeting or podcast, in one single pass. The final transcript must be detailed, showing who said what and when a line was spoken. It's also really important that I can provide a list of custom "hotwords," like technical terms or specific names, to make the recognition more accurate. This should support a bunch of different languages, not just English.

For the text-to-speech feature, it should create natural-sounding audio from long scripts. I want it to be able to generate a conversation involving a few distinct voices. A real-time streaming mode, where the app generates audio as you type, would be a great feature too. Use a modern Python framework for this and feel free to use powerful open-source models from places like Hugging Face.