direct speech reconstruction