Social recommendation using speech recognition: Sharing TV scenes in social networks

We describe a novel system which simplifies recommendation of video scenes in social networks, thereby attracting a new audience for existing video portals. Users can select interesting quotes from a speech recognition transcript, and share the corresponding video scene with their social circle with minimal effort. The system has been designed in close cooperation with the largest German public broadcaster (ARD), and was deployed at the broadcaster's public video portal. A twofold adaptation strategy adapts our speech recognition system to the given use case. First, a database of speaker-adapted acoustic models for the most important speakers in the corpus is created. We use spectral speaker identification for detecting whether one of these speakers is speaking, and select the corresponding model accordingly. Second, we apply language model adaptation by exploiting prior knowledge about the video category.