Recurrent neural network based Music Recognition using Audio Fingerprinting

摘要

In places like restaurants and shopping malls, the music that plays might sometimes captivate the customers. The aim of this research is to obtain an efficient aid for the people to discover the song and get acquainted with it. For this purpose, a sample input is fed into the audio fingerprinting and matching module which is fingerprinted just like the songs in the database. A simple linear search is performed against the database which contains the fingerprints of several songs. This algorithm gives ideal results for an input duration of 5 seconds and is scalable. The cover song identification is achieved by extracting the features like chroma features, spectral centroid, spectral contrast, MFCC and tempo from the original songs as well as its cover songs. Proposed work is implemented using Long Short Term Memory (LS TM), to provide the original song as the output based on a score which is the number of matching features from both the original song and its cover version. Singing detection is accomplished by performing lyrical matching. A live recording of a person's singing is translated into text. This translated input is linearly searched with the dataset containing the lyrics and the output is the corresponding song name.