Detection of Publicity Mentions in Broadcast Radio: Preliminary Results.

The advertising mentions are publicity contents that are not prerecorded, usually are said by radio or TV broadcasters to publicize a product or a company. The main difficulty of detecting advertising mentions is that the audio is not exactly repeated every time, as happens with conventional prerecorded advertising where more efficient techniques such as audio fingerprinting can be used. This paper proposes the use of a keyword search system in Spanish for the detection of advertising mentions. For that, it has been necessary to train and evaluate a new speech recognizer in Spanish (LVCSR) using the Kaldi tool and databases Fisher Spanish and Callhome Spanish. The best word error rate we have obtained on conversational telephone speech is 41.10 %. For the evaluation of mentions detection a specific database in Spanish has been created, containing 300 h of audio, 25 of which have been tagged with different types of information, including mentions appearing in the audio. The recognizer has been applied to all advertising mentions in search for mention specific keywords, achieving a detection rate of about 74 %.