压缩域鲁棒音乐指纹算法研究

对互联网海量MP3格式音乐数据进行基于内容的有效检索是当前一个重要而又很少涉及的研究方向.本文提出一种基于MDCT频谱熵的压缩域音频指纹算法，对各种常规频域和时间域的音频信号处理失真具有较强的鲁棒性.模拟实验在包含100首不同中文流行歌曲的音乐数据库上进行.对经受各种严重信号处理失真的粒度为5s左右的查询片段，能够取得超过90%的首位正确识别率.

A Robust Compressed-Domain Music Fingerprinting Technique Based on MDCT Spectral Entropy

With the proliferation of MP3 music,compressed-domain music information retrieval from the Internet has come into being an important and urgent research field.In this paper,we propose a novel compressed-domain audio fingerprinting algorithm based on MDCT spectral entropy.The input MP3 music file is first partially decompressed to obtain MDCT coefficients as intermediate results，whereby we calculate the MDCT spectral entropy through consecutive long windows and come to the final fingerprint sequence by magnitude relationship modeling.Such fingerprint exhibits strong robustness against various frequency-and time-domain audio distortions due to its statistically stable nature.Experimental results show that in our test database which is composed of 100 distinct Chinese pop songs，a 5s music clip is sufficient to identify its original recording in real time，with more than 90% top one precision rate even under various severe audio signal distortions.