Automated Content Based Audio Monitoring Approach for Radio Broadcasting

Abstract

Automated content based audio monitoring approach for radio broadcasting is a very important problem in world of multimedia computing. We propose a real time system to solve this problem which includes our own audio recognition algorithm. It is easy to recognize a song, when you provide the original high quality blueprint of the song as input. But we can t expect such kind of audio input from radio channels since lots of transformations are possible before reaching the end user or the listener. For example, adding environmental effects such as noise, adding of commercials, watermarks etc. These transformations cause change in the uniqueness of particular song and make the problem even more difficult. The algorithm we proposed is resistant to noise and distortion as well as it is capable of recognizing short segment of song when they are broadcasting. We extract sets of spectrum peaks of a given audio clip. These peaks can survive in a noise channel since peaks always spread around the zero axis. Because of this nature original peaks are not altered at the most of the time. Then we generate a unique audio fingerprint using these peaks. In order to deal with audio distortions, we split the song into sets of frames. Then we generate sets of fingerprints for every frame. Even though a part of a song is completely lost or distorted, we can use frames of the remaining part to identify the song correctly. Another usage of this framing approach is that we can filter only song frames for further processing. All non-song frames such as commercials, dramas etc. are discarded. The approach we proposed woks in real time. To achieve this, we use many threads. One thread is responsible for continuous buffering and splitting the buffered stream into equal size chunks. Another thread processes these chunks at the same time. At the end of the processing stage our system generates a descriptive report to the end user including all details of the broadcasted history. We use ten scenarios which occur in an actual radio channel. Then we evaluate the system against these real world scenarios and show the level of accuracy at the end. We obtained over 90% to 96% accuracy level for each case.