Symbolic speaker adaptation for pronunciation modeling