21 of 22 people found the following review helpful:
5.0 out of 5 stars
The beginnings of the machine musicians., February 3, 2004
This review is from: Machine Musicianship (Hardcover)
The author introduces this book as "exploration of the theoretical foundations of analyzing, performing, and composing music with computers". The book is exceptional in quality of writing and is fascinating to read. The fact that machines are actually composing music that is enjoyable to listen to as well as non-trivial from the judgment of professional musicians has to rank as one of the most fascinating achievements in music theory and artificial intelligence. Readers of this book will require a fair amount of background in music theory, some knowledge of mathematics (the fast Fourier transform is used in some parts of the book), some knowledge of various algorithms in artificial intelligence, such as neural networks and genetic algorithms, and knowledge of the C++ programming language. There is a CD-ROM that accompanies the book, and which contains a library of C++ objects that can be used to build interactive programs. The author emphasizes though the reader does not have to become a C++ programmer in order to follow the text. In addition, he is concerned with the ability and occurrence of machine musicians, and not in the question as to whether or not the methods they use can emulate human music cognition.
In chapter 1 the author introduces techniques necessary for the analysis of algorithms and composition, and for studying pitch-specific processes such as chord classification and key induction. The author restricts himself to what he calls "symbolic processes" in this chapter, and these he defines as those that are best characterized as a system of representations and rules. The symbols are taken to represent features of the musical context and their relationships are inferred by algorithms based on the knowledge of the objects they represent in real music. A context-independent chord classifier is first developed and then the author gradually introduces elements of context dependence with the goal of showing how these elements improve performance. The author also addresses issues with music representation, such as MIDI, and some possible successors to it. A key induction algorithm based on parallel processing is discussed in some detail. This algorithm takes knowledge about scales and chord functions and then updates saliency ratings for major and minor tonalities.
In chapter 2, the author deals with "sub-symbolic" processes, which he characterizes as processes that utilize the regularities learned from prior inputs in order to characterize and predict future inputs. He could have designated these as "connectionist" algorithms, since they learn their behavior from being exposed to material, do not depend on fixed rules, and are usually implemented using neural networks. Neural networks are used to do key induction, and the author discusses a connectionist approach to perform quantization, and a modification of it for use in real time. He also discusses various techniques for doing "beat tracking", i.e. the process of finding a regular pulse in a sequence of events, remarking that such an ability is very difficult to accomplish in a machine. Also discussed are some algorithms for performing meter induction.
The author turns his attention to pattern recognition and segmentation of music in chapter 4. The discussion of grouping preference rules leads the author to a real-time segmenter, whose rules are grouped according to the quantity of information that is required to apply them. The author discusses the problems with doing real-time segmentation. A dynamic programming approach to pattern matching is outlined, one using a "rating matrix", and the author points out that the difficulty in doing pattern matching does not lie in the matching algorithm, but rather in the preparation of the patterns needed by the algorithm. The "absolute" representation of the pitch used in this algorithm is then replaced by the "intervallic" representation in order to adhere to what is known from human music cognition. The author then compares his pattern matching algorithms with what has been done in the literature. All of this discussion is fascinating, particularly the discussion of Kohonen self-organizing neural nets that learn to cluster inputs into categories via competitive learning.
Then, in chapter 5, the author begins a discussion of the techniques that machines use to do musical composition. All of the algorithms used by these machines are interactive, in that they change their behavior in response to external inputs. Many different techniques are overviewed, such as generation techniques, which the author is well known for, Also discussed are score following, a pattern matching technique that can trace the progress of live performers through their compositions they are playing, and algorithmic signal processing. Detailed diagrams are given for illustrating the different algorithms.
This is followed in chapter 6 by a review of proposals for algorithmic performance and expression, emphasizing the role in particular of research done in music cognition. The author is careful in pointing out that he remains neutral concerning the question of whether a program is capable of musical cognition similar to what humans do. In addition, he gives two reasons why the experimental data from music cognition is not the standard of verification for the algorithms given in the book. Clearly the author is leaning towards the view, and I believe correctly so, that machine intelligence, even in contexts outside of music, may be in many ways very different from human intelligence. Machines will compose and produce music using techniques that may be very different from what humans use, just as machines play chess in ways that are very different from what humans do. Expert systems are discussed as a tool for algorithmic composition, and the author addresses issues of knowledge representation, with detailed emphasis on schemata for this purpose. Particularly interesting is the fact, as the author points out, that most interactive music systems do not learn, but the author discusses various methodologies that are attempting to incorporate learning into machine musicianship, such as neural networks and genetic algorithms. The GenJam machine of John Biles for doing jazz improvisation is a very interesting example of the latter.
I did not read chapters 7 or 8 so I will omit their review.
Help other customers find the most helpful reviews
Was this review helpful to you? Yes
No
