Thank you for the feedback. I was aware of the existence of the marginalized version of a GMM, but I had the idea that in that scheme, you loose information about the categorical variable z and you cannot do inference about the probability of a given point to be in a specific cluster. Is this idea correct?