Audio-based age and gender identification to enhance the recommendation of TV content

Recommending TV content to groups of viewers is best carried out when relevant information such as the demographics of the group is available. However, it can be difficult and time consuming to extract information for every user in the group. This paper shows how an audio analysis of the age and gender of a group of users watching the TV can be used for recommending a sequence of N short TV content items for the group. First, a state of the art audio-based classifier determines the age and gender of each user in an M-user group and creates a group profile. A genetic recommender algorithm then selects for each user in the profile, a single personalized multimedia item for viewing. When the number of items to be presented is different to the number of viewers in the group, i.e. M = N, a novel adaptation algorithm is proposed that first converts the M-user group profile to an N-slot content profile, thus ensuring that items are proportionally allocated to users with respect to their demographic categorization. The proposed system is compared to an ideal system where the group demographics are provided explicitly. Results using real speaker utterances show that, in spite of the inaccuracies of state-of-the-art age-and-gender detection systems, the proposed system has a significant ability to predict an item with a matching age and gender category. User studies were conducted where subjects were asked to rate a sequence of advertisements, where half of the advertisements were randomly selected, and the other half were selected using the audio-derived demographics. The recommended advertisements received a significant higher median rating of 7.75, as opposed to 4.25 for the randomly selected advertisements.