Seeking for best practice in Chinese multi-speaker scenario.

[Huowuge]

I'd like to seek advice on how to better utilize F5-TTS's pre-trained model for inference, primarily discussing scenarios with multiple speakers. I have tried using several 10-second Chinese voice recordings of myself and my friends. Even though I have made considerable efforts to read in standard Mandarin and carefully left about 1 second of silence at the end, the results are still unsatisfactory.

The main issues include unnatural connections, word segmentation errors, inconsistent volume, and inexplicable speed variations.

My questions are:

1. Should I perform more in-depth preprocessing on the input ref_audio and ref_text, such as using specific corpus content, controlling reading speed, and managing volume techniques?
2. Should I adjust model parameters for speed and volume instead of using default values?
3. Should I perform better preprocessing on gen_text, such as word segmentation and paragraph division?
4. Is the audio quality of .m4a recordings on an iPhone sufficient for zero-shot TTS?
5. Apart from these methods I've speculated, is fine-tuning necessary?

I would appreciate if anyone could answer my questions and share your best practices for multi-speaker scenarios. Thank you.