Cover song identification with direct chroma feature extraction from AAC files
Tai-Ming Chang, En-Ting Chen, Chia-Bin Hsieh, Pao‐Chi Chang · 2013
This paper proposes a low-complexity and effective feature extraction method derived directly from AAC files. Unlike traditional methods that must decode audio files and then compute fast Fourier transform coefficients, the proposed system directly maps the modified discrete cosine transform coefficients into a 12-dimensional chroma feature without fully decoding it. To accelerate the matching time, segmentation is applied to reduce the time dimension in the feature space. In addition, the dynamic programming technique is used to match songs to various tempos. The experimental results show that the proposed system achieves a 62% accuracy rate, which is an improvement over the traditional FFT-based system, and reduces the computational complexity by approximately 35%.