Odd results in ASR. Does it have a chat language model and text smoothing?

[cageyoko]

There are three cases that I found obvious different with reference labels.

1. When I said "hello", the output of model is "hello, how are you"

2. When I repeat said a same word, the output of model will delete the repeat part, and some filled pause such as "um" also will be deleted. I think it

3. Here are some output that's completely unrelated to audio.

I think it can be explained by outputting correct speech sentences based on semantic analysis instead of the original ASR results？