Specifying reference text in GPT-SoVITS leads to messy audio output

Hi there!

When playing around with your GPT-SoVITS fork I found that specifying reference text as the second argument to `get_tts_wav` function here:

https://github.com/alphacep/vosk-tts/blob/89b23a8b033133e25e3e7f53d07939645b8ea51c/training/gpt-sovits/inference_cli.py#L273

results in the output audio being either gibberish or completely silent.

Without specifying the reference text inference works as intended.

Is this an expected behavoiur?