Wolf Blitzer: Are They Saying Any- Thing about How the Fire May Have Started at Least so Far, Dan? Dan Simon: Not Yet, Wolf. Other than to Say That It Looks like It Started in the Back of the Vehicle

elevision networks produce a tremendous amount of information every day. Identifying who is speaking at an instant can help the structural organization of news content, so we can answer questions such as “what did person X say about an issue?” In the past, people recognition was studied by first training a system using a set of faces and audio with labeled names. Here, we propose a method to identify who is speaking throughout a video without providing external information or training (see the “Related Work in Speaker Identification” sidebar). We search for possible speaker names within the three components of TV news: video, audio, and closed captions. First, our system checks whether the name of a speaker appears on screen while the person speaks. It then determines the most common name to appear on screen during a speaker’s audio segments. Finally, our system parses the closed captions to see if the name of a speaker is used. By using weighted majority voting to combine this information, we produce a text transcript with speaker names. As a proof of concept, we focus on CNN news because it is more consistent than other networks in introducing speakers and displaying the speaker’s name on screen. One of the goals of our system is to label speakers in closed captions, because speaker transitions are marked with a “ ” prompt, but no speaker name is displayed except for prerecorded TV shows. We have added the underlined names of speakers (which are not part of the script) to a sample of closed captions from the “CNN Newsroom,” which aired 5 May 2013, at 12:00 pm, as follows: