A new approach to create high level features from low level features of audio clips

Existing search engines utilize text representing the required audio clips as the index for searching and retrieving the same from the huge database. Instead, a six dimensional high level feature vector based on the human feelings experienced on listening to the audio clips can be used as the index for searching, which is comparatively better in obtaining the optimal audio clips since it takes into account the user's subjective perception of similarity. The high level feature vector is defined as the vector consisting of the ranks assigned by humans based on the emotions experienced while listening to a song, such as happy, sad etc. However, it is not practical to assign a human ranked high level feature vector for the millions of songs available in the database. Hence, in this paper, we propose a technique of mapping the low level features of a song such as cepstrum and harmonicity to its high level features.