Evaluating the Suitability of SoundFingerprinting for Music Identification in Movies

Abstract

Audio fingerprinting has shown to be an effective approach to music identification, having properties robust to noise and signal degradations. A field in which audio fingerprinting has not been evaluated yet is music identification in movies. In movies, music is often accompanied with background noise, sound effects and dialogue, and further processed using mixing and mastering techniques. This paper evaluates the suitability of the audio fingerprinting framework `SoundFingerprinting' for the identification of music in movies. The framework is evaluated according to a benchmark established for this field. The framework was tested on actual movie data, noise-layered soundtracks, pitch-shifted soundtracks and tempo-changed soundtracks. The framework was unable to identify the music in actual movie data, thus directing the research to identify problematic areas specific to SoundFingerprinting. In identifying noise-layered soundtracks, the framework showed varying performance dependent on the dominant frequencies present in the noise sample. Furthermore, the framework showed to be robust to tempo-changes, whereas the framework was unable to identify pitch-shifted soundtracks. Based on this performance evaluation, SoundFingerprinting is ill-equipped for the task of music identification in movies.