Novel Algorithmic Approaches & Machine Learning for Analysis of Musical Signals

这项研究的目的是研究、设计和开发新的基于算法和机器学习的实时音乐信号分析方法。本研究旨在突破现有知识和方法的局限，提出一种新的方法来解决该领域尚未解决的问题，如复调基音检测。作为计算机科学、电子工程和音乐等多种学科的交叉，这一领域的研究也让位于各种跨学校合作的机会，以及帮助扩展一系列不同领域的知识。音乐一直是人类文化的基本组成部分--已知的最早的乐器可以追溯到大约36000年前(Morley，2003年)。作为生活中如此根深蒂固的一面，有太多的音乐理论和数学支撑着整个领域。随着计算机在上个世纪的出现，音乐信号的分析是一个日益增长的研究领域，它面临着一系列具有挑战性的问题，例如基音检测(单音和复音)、盲源分离和现场合奏中的音乐同步等等。本研究的目的是开发有效和准确的方法来解决这一领域的问题，以及对各种方法解决这些问题的可行性进行调查。此外，探索解决单一问题的多种方法并在一系列环境中比较和对比这些实验的结果是令人难以置信的有趣的。首先，将对当前和历史方法进行深入调查，并研究学者和商业“用户”对新方法的感知要求。然后，这些调查的数据可以用来确定需要更可靠或更有效的方法才能有效的关键分析。对于每一个已确定的领域，必须开发和测试多种方法(包括新颖的和基于机器学习的)，以确定哪些方法在某些条件下或在不同的环境中执行得最好。然后将迭代地改进其中每一项的最佳表现，以期在确定基准时达到预定的精度和效率。由于这项研究是计算机科学和电气工程之间的联合合作，比较和对比论文组中其他人对该领域类似问题所采取的不同方法尤为重要，特别是因为其他人可能有不同的方法，很容易影响或改进当前的实现。

The aim of this research is to investigate, design and develop both novel algorithmic and machine learning based approaches to the analysis of musical signals in real-time. This research aims to push the boundaries of current knowledgeand approaches to this kind of analysis, proposing new methods of tackling yet-unsolved problems in the field such as polyphonic pitch detection. As an intersection between such a wide variety of disciplines including computer science, electrical engineering and music, research in this space also gives way to various opportunities for cross-school collaboration as well as helping to expand knowledge in a range of distinct fields.Music has been a fundamental part of human culture for aeons - with the earliest known instruments found to date back c. 36000 years (Morley 2003). As such a deep-rooted facet of life, there is a plethora of musical theory as well asmathematics that underpins the field as a whole. With the emergence of computers in the last century, analysis of musical signals is an increasingly growing area of study that is host to a range of challenging problems, forexample pitch detection (both monophonic and polyphonic), blind source separation and musical synchronisation in live ensembles to name but a few.The purpose of this research is to develop efficient and accurate approaches to problems in this space, as well as performing investigations into the viability of various approaches to these problems. Moreover, it is incredibly interesting toexplore multiple approaches to single problems and compare and contrast the results of these experiments over a range of environments.Initially, an in-depth investigation into current and historical approaches will be undertaken, as well as a study into the perceived requirements of new approaches as seen by both academics and commercial "users". The data from these investigations can then be used to identify key pieces of analysis that require more reliable or efficient methods in order to be effective. For each of these identified areas it will be imperative to develop and test multiple approaches (both novel and machine learning based) in order to ascertain which methods perform best under certain conditions or in different environments. The best performing of each of these will then be iteratively improved with the aim of reaching a predetermined accuracy and efficiency when bench marked. As the research is a joint collaboration between computer science and electrical engineering, it is especially important to compare and contrast the different approaches that others in the thesis group have taken to similar problems in the space, especially because others are likely to have varying approaches that could easily influence or improve current implementations.