A study on music annotations based on the compressed domain of MPEG audio

在这项研究中，我们集中于使用在TwinVQ压缩的编码步骤中计算的序列或从压缩音乐数据中的比特流来进行音乐标注。我们提出了三种方法，1)利用自相关系数序列进行音乐结构分析，2)利用码流计算的LPC倒谱序列对音乐类型进行分类，3)直接从码流中提取LSP参数进行印象词标注。

In this study, we focused on music annotation using a sequence computed in encoding step of TwinVQ compression or from a bit-stream in compressed music data. We proposed three methods for, 1) music structure analysis using a sequence of auto-correlation coefficients, 2) music genre classification using a sequence of LPC cepstrum computed from bit-stream, 3) impression word annotation using LSP parameter directly extracted from bit-stream.