Large Scale Similar Song Retrieval using Beat-aligned Chroma Patch Codebook with Location Verification

With the popularity of song search applications on Internet and mobile phone, large scale similar song search has been attracting more and more attention in recent years. Similar songs are created by altering the volume levels, timing, amplification, or layering other songs on top of an original song. Given the large scale of songs uploaded on the Internet, it is demanding but challenging to identify these similar songs in a timely manner. Recently, some state-of-the-art large scale music retrieval approaches represent songs with a bag of audio words by quantizing local features, such as beat-chroma patches, solely in the feature space. However, feature quantization reduces the discriminative power of local features, which causes many false audio words matches. In addition, the location clues among audio words in a song is usually ignored or exploited for full location verification, which is computationally expensive. In this paper, we focus on similar song retrieval, and propose to utilize beat-aligned chroma patches for large scale similar song retrieval and apply location coding scheme to encode the location relationships among beat-aligned chroma patches in a song. Our approach is both efficient and effective to discover true matches of beat chroma patches between songs with low computational cost. Experiments in similar songs search on a large song database reveal the promising results of our approach.