Speech quality estimation has recently undergone a paradigm shift from human- hearing expert designs to machine-learning models. However, current models rely mainly on supervised learning, which is time-consuming and expensive for label collection. …