Context-sensitive evaluation of automatic speech recognition