add text condition for gen music

[jiaqianjing]

The music cannot be controlled through text. It's better with text constraint like text2img, I think that.

[BlinkDL]

it's possible by using such data to train / finetune the model