Generation of pronunciations from orthographies using transformation-based error-driven learning