Learning musical structure with machines

使用标准的时间序列分析技术很难或不可能学习音乐序列。这是由于音乐结构是在很长的时间跨度上精心制作的，并且是分层的，这使得很难或不可能从本地信息中学习。通过示例将短语“John ate the >”与音乐序列“B C E >”进行比较，在语言学情况下，单词“John ate the”约束了下一个单词>的可能值。 在音乐的例子中，附近的音符“B”、“C”和“E”对接下来的内容几乎没有什么约束。 这并不是说音乐缺乏约束。 在音乐中，这些限制来自于音乐结构背景下的精心制作和重复。 不幸的是（至少对于标准的序列学习工具来说），这种结构是在很长的时间跨度上，在不同的时间尺度上并行进行的。我提出了一个调查音乐是如何在时间上的结构，以及音乐结构如何可以用来约束机器学习算法应用在音乐领域。这项研究的双重目标是：（a）进一步了解音乐是如何结构化的，以及音乐结构如何与音乐感知和制作相关;（B）改进机器学习算法的最新技术，以实现两个重要的音乐相关任务：人类逼真的音乐表演生成和音乐序列学习。我还将解决相关的任务，如音乐分类，音乐相似性测量和在线处理方法与真实的音乐家合作。所有这些研究背后的关键思想是，更好地理解音乐是如何及时构建的，将直接导致更好的机器学习算法用于音乐相关任务。这项工作应该会对机器学习、音频信号处理和音乐认知产生影响。本研究的部分内容应与处理语音、视频和相关时间序列的领域相关。

Music sequences are difficult or impossible to learn using standard time-series analysis techniques. This is due to the fact that musical structure is elaborated over long time spans and is hierarchical, making it difficult or impossible to learn from local information. By example compare the phrase "John ate the >" to the musical sequence "B C E >" In the linguistic case, the words "John ate the" constrain the possible values of next word >. In the musical case, the nearby notes "B" "C" and "E" do little to constrain what can come next. This is not to say that music lacks constraints. In music the constraints come from elaboration and repetition within the context of musical structure. Unfortunately (for standard sequence learning tools, at least) this structure is elaborated over long time spans and at different timescales in parallel. I propose an investigation of how music is structured in time and how music structure can be used to constrain machine learning algorithms applied in the domain of music. The two-fold goal of the research is to (a) further our understanding of how music is structured and how music structure relates to music perception and production and to (b) improve the state-of-the-art for machine learning algorithms for two important music-related tasks: human-realistic music performance generation and music sequence learning. I will also address related tasks such as music classification, music similarity measurement and online processing methods for collaboration with real musicians. The key idea behind all of this research is that a better understanding of how music is structured in time will lead directly to better machine learning algorithms for music-related tasks. This work should have impact in machine learning, audio signal processing and music cognition. Portions of this research should be relevant in domains dealing with speech, video and related time series.