QwenLM/Qwen3-TTS ? reverse-engineered prompt

I want to build a really good text-to-speech application. The main thing is that the voices should sound super natural and expressive, not robotic. It needs to handle a bunch of languages like Chinese, Japanese, English, and Spanish.

The app should have two killer features. The first is voice cloning from a very short audio sample – like, take 3 seconds of a voice and then be able to generate any text in that voice. The second is 'voice design,' where I can just describe a voice using words, for example, "a cheerful young girl's voice" or "a deep, serious narrator," and it will create it.

It should also be fast enough to do this in real-time for streaming. Let's wrap this all up in a simple web UI where you can type in text, choose or create a voice, and hit play.