A Study on Music Information Retrieval using Compressed Domain of MPEG-4

我们本研究的目的是使用 TwinVQ 音频压缩编码步骤中计算的系数构建一个基于内容的音乐检索系统。今年，我们重点研究基于内容的音乐推荐系统和音乐检索系统，该系统允许我们通过所覆盖的音乐进行查询。为了构建音乐推荐系统，我们提出了一个基于内容的音乐推荐系统的框架。在我们的框架上，我们构建了音乐检索功能，允许我们按节奏和音色进行查询。这些函数基于 TwinVQ 音频压缩编码步骤中计算的自相关系数和 LSP 倒谱。然后，我们利用音乐检索功能构建了基于内容的音乐推荐系统。实验结果表明，我们的基于内容的音乐推荐系统是有效的。对于允许我们通过所覆盖的音乐进行查询的音乐检索功能，我们重点关注在TwinVQ的编码步骤中计算的自相关系数。我们评估了自相关系数相对于音高和节奏的鲁棒性。实验结果表明了对于音调的鲁棒性。

Our purpose of this study is to construct a content-based music retrieval system using coefficients computed in the encoding step of TwinVQ audio compression. In this year, we focus on a content-based music recommendation system and music retrieval system which allows us queries by the covered piece of music.For constructing music recommendation system, we proposed a framework of content-based music recommendation system. On our framework, we constructed music retrieval functions which allowed us queries by rhythm and timbre. These functions are based on the autocorrelation coefficients and LSP-Cepstrum computed in the encoding step of TwinVQ audio compression. Then, we constructed content-based music recommendation system using the music retrieval functions. Experimental results indicated that our content-based music recommendation system is effective.For music retrieval function which allows us queries by the covered piece of music, we focus on the autocorrelation coefficients computed in the encoding step of TwinVQ. We evaluated robustness of autocorrelation coefficients with respect to pitch and tempo. Experimental results indicated the robustness with respect to pitch.