Only consider linear frames for talk detection -- ignore anything else (bug #2219)