Same news is good news: automatically collecting reoccurring radio news stories

We present methods for finding same or almost same news stories in the hourly radio news broadcasts. Our procedures are able to detect reoccuring news stories of subsequent news broadcasts spoken by the same or different announcers only from the speech signal. They allow to establish a large database of repeated and professionally read speech at low costs that is especially interesting for prosody research, but also, e.g., for concept-to-speech and socio-linguistic studies. An automatically recorded complete radio news broadcast is first segmented into individual news stories using HMM recognition. Then, the word sequence estimates of the stories are either compared directly (naive method) or realigned with the signal of other stories (realignment method) in order to find out which stories were read before and which not. Both methods can be further improved by computing “meta distances” that also take into account distances to other stories. We evaluate and compare the usefulness of the proposed methods on real life data. We find that the realignment method combined with meta distances is the most reliable of the methods and that it is well suited for the task.