We propose a system, the Continuator, that bridges the gap between two classes of traditionally incompatible musical systems: 1) interactive musical systems, limited in their ability to generate stylistically consistent material, and 2) music imitation systems, which are fundamentally not interactive. Our purpose is to allow musicians to extend their technical ability with stylistically consistent, automatically learnt material. This goal requires the ability for the system to build operational representations of musical styles in a real time context. Our approach is based on a Markov model of musical styles augmented to account for musical issues such as management of rhythm, beat, harmony, and imprecision. The resulting system is able to learn and generate music in any style, either in standalone mode, as continuations of musician's input, or as interactive improvisation back up. Lastly, the very design of the system makes possible new modes of musical collaborative playing. We describe the architecture, implementation issues and experimentations conducted with the system in several real world contexts.

This paper discusses the use of statistical models for the problem of musical style imitation. Statistical models are created from extant pieces in a stylistic corpus, and have an objective goal which is to accurately classify new pieces. The process of music generation is equated with the problem of sampling from a statistical model. In principle there is no need to make the classical distinction between analytic and synthetic models of music. This paper presents several methods for sampling from an analytic statistical model, and proposes a new approach that maintains the intra opus pattern repetition within an extant piece. A major component of creativity is the adaptation of extant art works, and this is also an efficient way to sample pieces from complex statistical models.

In this paper, we describe and compare two methods for unsupervised learning of musical style, both of which perform analyses of musical sequences and then compute a model from which new interpretations / improvisations close to the original's style can be generated. In both cases, an important part of the musical structure is captured, including rhythm, melodic contour, and polyphonic relationships. The first method is a drastic improvement of the Incremental Parsing (IP) method, a method derived from compression theory and proven useful in the musical domain. The second one is an application to music of Prediction Suffix Trees (PST), a learning technique initially developed for statistical modeling of complex sequences with applications in linguistics and biology. 1 Style Modeling By Style Modeling, we imply building a computational representation of the musical surface that captures important stylistic features hidden in the way patterns of rhythm, melody, harmony and polyphonic r...

The Continuator system is an attempt to bridge the gap between two classes of traditionally incompatible musical systems: 1) interactive musical systems, limited in their ability to generate stylistically consistent material, and 2) music composition systems, which are fundamentally not interactive. The purpose of Continuator is to extend the technical ability of musicians with stylistically consistent, automatically learnt musical material. This requires the ability for the system to build operational representations of musical styles in real time, and to adapt quickly to external musical information. The Continuator is based on a Markov model of musical styles augmented to account for efficient real time learning of musical styles and to arbitrary external bias. The paper describes the main technical issues at stake concerning the integration of an agnostic learning scheme in an interactive instrument, and reports on real-world experiments performed with various musicians.

We describe variable markov models we have used for statistical learning of musical sequences, then we present the factor oracle, a data structure proposed by Crochemore & al for string matching. We show the relation between this structure and the previous models and indicate how it can be adapted for learning musical sequences and generating improvisations in a real-time context.

We studythe synchronization of musical sequences bymeans of an operation de ned on nite or in nite words called superimposition. This operation can formalize basic musical structures such as melodic canons and serial counterpoint. In the case of circular canons, we introduce the superimposition of in nite words, and we present an enumeration algorithm involving Lyndon words, which appear to be a useful tool for enumerating periodic musical structures. We also de ne the superimposition of nite words, the superimposition of languages, and the iterated superimposition of a language, which is applied to the studyof basic aspects of serial music. This leads to the studyof closure properties of rational languages of nite words under superimposition and iterated superimposition. The rationalityof the transduction associated with the superimposition appears to be a powerful argument in the proof of these properties. Since the superimposition of nite words is the max operation of a sup-semilattice, the last section addresses the link between the rationalityof a sup-semilattice operation and the rationalityof the order relation associated with it. c ○ 2003 Elsevier B.V. All rights reserved.

Abstract. The paper concerns the use of multiple viewpoint representation schemes for prediction with statistical models of monophonic music. We present an experimental comparison of the performance of two techniques for combining predictions within the multiple viewpoint framework. The results demonstrate that a new technique based on a weighted geometric mean outperforms existing techniques. This finding is discussed in terms of previous research in machine learning. 1

Markov chains are a well known tool to model temporal properties of many phenomena, from text structure to fluctuations in economics. Because they are easy to generate, Markovian sequences, i.e. temporal sequences having the Markov property, are also used for content generation applications such as text or music generation that imitate a given style. However, Markov sequences are traditionally generated using greedy, left-to-right algorithms. While this approach is computationally cheap, it is fundamentally unsuited for interactive control. This paper addresses the issue of generating steerable Markovian sequences. We target interactive applications such as games, in which users want to control, through simple input devices, the way the system generates a Markovian sequence, such as a text, a musical sequence or a drawing. To this aim, we propose to revisit Markov sequence generation as a branch and bound constraint satisfaction problem (CSP). We propose a CSP formulation of the basic Markovian hypothesis as elementary Markov Constraints (EMC). We propose algorithms that achieve domain-consistency for the propagators of EMCs, in an event-based implementation of CSP. We show how EMCs can be combined to estimate the global Markovian probability of a whole sequence, and accommodate for different species of Markov generation such as fixed order, variable-order, or smoothing. Such a formulation, although more costly than traditional greedy generation algorithms, yields the immense advantage of being naturally steerable, since control specifications can be represented by arbitrary additional constraints, without any modification of the generation algorithm. We illustrate our approach on simple yet combinatorial chord sequence and melody generation problems and give some performance results.

The study is based on my dissertation proposal. Its purpose is to improve my understanding of the feature extractors used in the field of content-based music retrieval/classification. I am particularly interested in Jonathan Foote’s self-similarity method. I have summarized various articles related to feature extractors used in the field of audio information retrieval/classification. I have also included two experiments with Foote’s method on real musical pieces. Even though the analyses of these experiments are still in progress, this study has helped me in understanding the structural process of Foote's method. It seems that each variable in the system such as frame rate, kernel size, etc., should be optimized for each case by conducting thorough empirical experiments on different kinds of musical signals. Moreover, processing with this system on a full piece of music (usually more than 1 minute long) requires a lot of computing time. Hence, one of the final aim of my study is to suggest ways of improving some of the limitations of this system.

The fundamental problem investigated is to develop a model space that automatically classifies musics into categories by the detection of perceptually similar patterns in music. The concept is based on a modular system consisting of three main stages. The first stage involves the preprocessing of the raw audio data with passing over by a number of independent feature extractors. Each feature extractor reduces the information content in the raw music data to a vector in a small number of dimensions. In the second stage the set of feature vectors are classified(indexed) into certain clusters by pattern-matching algorithm. And in the final stage, the query engine detects/retrieves similar trajectories of an example in the database by means of a similarity-matching algorithm. In addition, several challenging issues will be addressed in this dissertation. Firstly, an exploration of ways to measure subjective perceptual qualities of musical signals. How do musical dimensions (e.g., pitch, rhythm, melody, tempo, harmony, etc.) perceptually integrate and interact in order to allow us to perceive a comparative sense of similarity between two sounds sharing particular attributes and structures. Secondly, the development of a good similarity measure for perceptually salient index attributes. Also, if the model space can be evaluated/improved with a human-classified ”ground truth ” dataset, it would result in a good psychological model of human perception and categorization of music. 2 1