ISE599: Engineering Approaches to Music Perception and Cognition

Week 13 (16 Apr 2003): Pattern Recognition



Summary of the third class
by Yun-Ching Chen, Cindy Lee, Andy Lee, Frances Kao, Arpi Mardirossian, Sean Mo, Xumei TAN, Erdem Unal, Phillip Wu, Shivani Yardi

Yun-Ching CHEN

In this lecture, we first experienced some examples of computer-generated music. Personally, I think these works are not too bad; they just missed some human-expressions. But, thatb in AI field; at least they did pretty well in the idea of generating music purely by computer.

After that, we had three presentations, which two of them are from David Cope who is doing the computer-generated music research, and the other one is from MIT media lab. The first one tried to represent the music pattern/style from linguistic aspect where the second one tried to distinguish them in signatures and earmarks. They are both ways to be a symbol of music pattern/style and able to help in building music. The third paper talked about the music style classification. Itbpic but they should have more several of samples in their study.

Cindy (Chia-Ying) LEE

This week we have three paper presentations about abstracting and recognizing styles in music. Two papers are from David Cope. The first one is "On algorithmic Representation of Musical Style", in which he thinks music is associative to language. Therefore he used linguistic models to represent the stylistic patterns in music. The model can represent the hierarchy of importance and function of each music events. He also used these representations to generate replications of masterpieces, though some results are not agreeable.

The second David Cope's work is "Signatures and Earmarks: Computer Recognition of Patterns in Music". He thinks music stylistic patterns should be distinguished into two type, signatures and earmarks. According to the context, signatures are common motives existing in two or more works of a composer; earmarks are patterns the composer likes to use in a certain structure location of some certain movements. For example, in the first movement of his piano concertos, Mozart liked to use a scale going to trills before cadenzas. But the definitions of these two patterns sound vague to me.

The third paper is "Combining Musical and Cultural Features for Intelligent Style Detection". The authors used neural network to classify music from different genres.

Andy (Ming-Chang) LEE

This lecture was organized by two presentations, which were given by Cindy, Andy and Frances respectively. In the beginning, Cindy talked about the Language Parsing (Hierarchy Parsing) approach for music modeling. She also mentioned about the background/foreground structure in music, and the Shenkerian layer analysis (tonic prolongation): every piece can be reduced to "do". Elaine and Shivani cooperatively gave us an example for the Shenkerian layer analysis. Then, Cindy introduced the Experiments in Musical Intelligence (EMI), which is a spec and also a program for music analysis and creation. After Cindy's presentation, I talked about two important musical patterns, signatures and earmarks. Signatures are specific styles belong to specific composers. With signature analysis, we can understand the time slot of a composer and we can also tell a music piece of one composer from other composer. Earmarks are everything else but not signature. For composers during same periods, earmarks could appear in anywhere in their works, like trill. Earmarks contains some important messages, to expect the climax, to connect two important parts, to link parts to Cadences, etc. Then, Frances gave a presentation about Style Detection. This presentation arouse arose lots discussions, since their approach and results are not persuasive, and people have their own classification for different music.

Frances (Hui-Yun) KAO

The paper presentations we had on that class are mainly about computer-generated music patterns and style classification. The first two papers are from David Cope, and the other one is from MIT Media Lab. We listened to some computer generated inventions, and frankly speaking, I think those inventions are good. What I mean by "good" is that they are "nice-listening", not that they are just like the originals. I think David Cope's work is successful at the aspect that his algorithm creates nice music. The MIT paper is interesting to me, but as Alex and Elaine mentioned, the approaches they propose have the problem with scaling up to large dataset.

Arpi MARDIROSSIAN

This class focused on musical style. We began by listening to a variety of Bach pieces in different styles, as well as some computer-generated pieces. We also had three presentations.

The first presentation by Cindy was on "On the Computer Recognition of Musical Style". This paper focuses on drawing parallels between language and music. Parsing methods that are used in language are expanded to music.

The second presentation by Andy was on "Signatures and Earmarks: Computer Recognition of Patterns in Music." This paper focuses on breaking down the elements that make up patterns in music. This is necessary in order to create algorithms with which a computer can detect patterns. Two such elements are signatures and earmarks.

The third presentation by Frances was on "Combining Musical and Cultural Features for Intelligent Style Detection." This paper proposes a system that can detect style based on acoustic content. A neural network is utilized for classification.

Sean (Zhenyao) MO

This week we had three presentations.

First is "On the Computer Recognition of Musical Style" by David Cope, which was presented by Cindy. It's about treating music as language, analyzing the structure, and decomposing the music piece into nodes and paths, forming the Augmented Transition Network (ATN). Once we have ATN, we can synthesize a new piece by parsing this network.

Second is also a David Cope paper "Signatures and Earmarks: Computer Recognition of Patterns in Music", presented by Andy. I learned that though both signatures and earmarks are about features of a piece of music, they are totally two different things. Signature is defined as the characteristics of a composer in music, whereas earmarks are about general music structures (not related to a specific author).

Third one is presented by Francis, "Combining Musical and Cultural Features for Intelligent Style Detection", a paper from B. Whitman and P. Smaragdis. This is about music classification. By analyzing the acoustic data, the music piece is decided to be a "country", a "Rap", a "R&B", or something else. But both the method and results are not persuading as to me.

Xumei TAN

At first we listened to several Bach style music pieces including fugue. I think it doesn't sound bad. And this kind of music is generated by computer.

Then we had three presentations talking about this research area. The first one is about the paper: "on algorithmic representation of musical style" written by David Cope. He use the linguistic rule to find the music pattern. He designed the EMI system. There are OSAC in three levels of importance. Dr. Chew then showed us that any piece in C major can be reduced to C. Then the author use this way to find the most essential pattern in a piece of music and by keeping that part and doing minor change in other parts, he can generate the music by computer.

The second presentation was also about David Cape's paper: "Signatures and Earmarks. He tries to distinguish between these two types of pattern. Signature is something unique for a composer. It can tell us what period of music history a work comes from and the probable composer of that work. It's normally just 2 or 5 beats in length and often composites of melodic, harmonic, and rhythmic elements. Signature usually occurs between 4 and 10 times in any given work. Variation often includes transposition, diatonic interval alteration, rhythmic refiguring, and registrar and voice shifting. This paper gives some example on how the signature varies but still can be recognized. It was used by EMI, allowing one step difference, different direction and extra note to recognize real signature. Earmark can be used by different composer and can show the structural events such as cadence.

The third paper was about "combining musical and cultural features for intelligent style detection". It described a computer based method to classify music styles. The system can operate on acoustic content of the audio and community metadata: a vector space of descriptive textual terms crawled from the web. It used the neural network to do classifying. The result is quite interesting. They choose different artist and map them to 5 styles. And the paper claims that if combine the audio based and community web based method they can get 100% accuracy on classification.

Erdem UNAL

In the beginning of the class we listened some music that is claimed to be recomposition(!) of Bach's masterpieces. And the lecture was more focused on this kind of music generation and basically two presentations from David Cope reviewed his work.

Using the Shenkerian analysis this guy was trying to reduce the contents of a piece and turn into a combination of basic tones. This may be an interesting idea, however as I was trying to claim in the class, he shouldn't be dealing with Bach's or other music master's work. It is obvious that this technique is loosing a lot of artistic information from the whole music piece and turns it into a robotic style. I was able to distinguish the Bach from its computer recompositions. Another thing, I didn't really understand what was the purpose of this research.

And another presentation talked about David Cope's ideas about signature and earmark research. He was trying to fin out some patterns in the music that may carry specific information about the music itself such as writer, its historical time etc...

The third paper was about combining musical and cultural features. In this paper, acoustic information was gathered from different kinds of music and it is claimed that this fatures can create classifications by the help of neural network. This paper was interesting...

Phillip WU

We listened to some pieces of computer-generated music which are the mimics of Bach pieces. The study is pretty meaningful because the researcher was trying to extract the features of those Bach's pieces.

We had 2 out of 3 paper presentations mentioning David Cope's work. He was using language parsing techniques and linguistic models to analyze the stylistic patterns of music. Here comes the hierarchy structure again. He used the hierarchy structure to map the importance and function of each musical event. The most interesting things are the concepts of Signature and Earmark he brought to us. I would like to describe the Signatures as the composer-oriented, while the Earmarks as the movement-oriented.

Shivani YARDI

In this class, we got an insight into various methods of style and pattern recognition by means of the presentations.

There was an analogy drawn between models built to explain linguistic structures and those used to detect musical structures.

The Shenkerain analysis was interesting as it can be broadly applied to almost any piece of music to collapse the complex musical structure around a few basic tones. We learnt of "signatures' and 'earmarks' as put forth by David Cope that are crucial in identifying the musical structure and style of a composer. Signatures, from what I understood are more 'motif' oriented, whereas 'earmarks' pertain to the musical structure in general. Signatures are found within diffrent works of a composer, and earmarks are patterns that could be found across works by different composers.


Posted 23 Apr 2003
The University of Southern California does not screen or control the content on this website and thus does not guarantee the accuracy, integrity, or quality of such content. All content on this website is provided by and is the sole responsibility of the person from which such content originated, and such content does not necessarily reflect the opinions of the University administration or the Board of Trustees