A novel singing inverse method is presented in this report. A complete pronunciation physical model is constructed by using the Two Mass model of vocal cord vibration and the Tube Resonance Model (TRM) of vocal tract. By employing the physical model, speech and singing voice can be synthesized by vocal cord and vocal tract physical parameters. A new self-supervised approach is presented to solve inverse problem. The proposed approach works is an analysis-by-synthesis way to learn an inference network by iterative sampling and training. The result has an explicit physical meanings. By this method, we can retrieve the physical parameters of vocal cords and tract of singers through the singing signal, so as to quantitatively analyze the technical characteristics of singers of different parts and singing styles.