A Deep Joint Sparse Non-negative Matrix Factorization Framework for Identifying the Common and Subject-specific Functional Units of Tongue Motion During Speech.