Generation inconsistencies

[samuelzxu]

Hi! Congratulations on the awesome product!

I tried generating the same prompts as from the demo, and was met with a few odd results that differed from the previous generated results. Using the same colab notebook, I got mostly silence for the spanish text, and some harsh screeching interspersed throughout the other prompts as well.

Here is the link to the colab notebook with the generated sounds: https://colab.research.google.com/drive/1iJtfgTCs3WgE0kfSQYEY1-XCy9G-TAt3#scrollTo=8KV3klnr-lvo

[gkucsko]

hm, so variation every time you generate is expected and kind of the cool thing about a model like this. it might create something that sounds like a political speech or two people arguing at a train station. Weird sounds can sometimes happen cause it's an autoregressive model. you can play with temperature to reduce the issue. might need different parameters for different languages and different history_prompts. Some of the prompt probably also work better than others