随着时代的发展和进步，音乐的创作水平和创作理念也应与时俱进。然而由于听众基数扩大及创作环境的改变等因素的影响，华语乐坛的歌曲创作水平却呈现出逐渐退步的趋势，其具体的表现为词曲质量降低、编曲单一、和弦老套等。

这些问题导致新歌并不能长久吸引听众。而各大音乐平台的热歌榜上，尽管新歌频繁更替，许多经典老歌却始终保留自己的一份位置。这一现象引起了一个假想：是否老歌的经久不衰源自于其不同于新歌的一些音乐特征？

基于上述假想，本文以构造一个多维音乐特征提取和量化系统为目标，以统计领域和音乐领域交叉为创新，结合音乐相关理论和频域分析技术，完成了以下工作：

第一，基于QQ音乐、网易云音乐、酷狗音乐和酷我音乐四大音乐平台的每日热歌榜完成歌曲下载后的伴奏人声分离、爬虫采集歌曲基本信息等工作。

第二，结合音乐理论和频域分析理论，将音乐特征的提取分为节奏特征、旋律特征、人声特征的提取以及风格分类四个子任务。其中，在旋律特征的提取过程中结合常数Q变换（CQT）以及基频识别等方法优化了旋律提取效果；在音乐风格分类过程中通过数据集重构训练出更适用于华语歌曲风格分类的SVM分类模型。

第三，以“是否为老歌”为因变量，对获取到的音乐特征变量进行二元Logistic回归。研究结果表明，华语经典老歌在维持歌曲律动感的同时，拥有更丰富和完整的歌曲结构、更密集且平滑变化的旋律以及更多的人声共鸣技巧的运用。这为华语乐坛未来的音乐创作与发展提供了方向上的启示。

With the development and progress of the times, the level and concept of music creation should also keep pace with the times. However, due to factors such as the expansion of the audience base and changes in the creative environment, the level of song creation in the Mandopop industry has shown a gradual decline, which is manifested by a decrease in the quality of lyrics and music, a single arrangement, and outdated chords.

These issues result in new songs not being able to attract audiences in the long term. On the popular song charts of major music platforms, despite frequent updates of new songs, many classic and old songs still retain their place. This phenomenon has led to an assumption: does the longevity of old songs stem from some musical features that are different from new songs?

Based on the above assumptions, this article aims to construct a multi-dimensional music feature extraction and quantification system, with the innovation of the intersection of statistical and music fields. Combining music related theories and frequency domain analysis techniques, the following work has been completed:

Firstly, based on the four major music platforms of QQ Music, NetEase Cloud Music, Kugou Music, and Kuwo Music, the daily popular song charts are completed, including the separation of accompaniment vocals and the collection of basic song information through web crawlers after song downloads.

Secondly, combining music theory and frequency domain analysis theory, the extraction of music features is divided into four sub tasks: rhythm feature extraction, melody feature extraction, vocal feature extraction, and style classification. Among them, the extraction process of melody features was optimized by combining constant Q-transform (CQT) and fundamental wave identification; In the process of music style classification, a SVM classification model that is more suitable for Chinese song style classification is trained through dataset reconstruction.

Thirdly, using "whether it is an old song" as the dependent variable, binary logistic regression was performed on the obtained music feature variables. The results indicate that classic Chinese old songs have a richer and more complete song structure, denser and smoother melodies, and more use of vocal resonance techniques while maintaining the rhythm of the songs. This provides directional inspiration for the future music creation and development in the Mandopop industry.