Identifying the common and subject-specific functional units of speech movements via a joint sparse non-negative matrix factorization framework