it sporadically does not generate the audio for the response, even though I am seeing that the model generate the text and that it was received correctly.