Search

Neural networks and recurrent neural networks have been employed to learn, generalize, and generate musical examples and pieces. Yet, these models typically suffer from an inability to characterize and reproduce the long-term dependencies of musical structure, resulting in products that seem to wander aimlessly. We describe and examine three novel hierarchical models that explicitly operate on multiple structural levels. A three layer model is presented, then a weighting policy is added with two different methods of control attempting to maximize global network learning. While the results do not have sufficient structure beyond the phrase or section level, they do evince autonomous generation of recognizable medium-level structures.

Throughout music history, theorists have identified and documented rules that capture the decisions of composers. This paper asks, "Can a machine behave like a music theorist?" It presents MUS-ROVER, a self-learning system for automatically discovering rules from symbolic music. MUS-ROVER performs feature learning via n-gram models to extract compositional rules — statistical patterns over the resulting features. We evaluate MUS-ROVER on Bach's (SATB) chorales, demonstrating that it can recover known rules, as well as identify new, characteristic patterns for further study. We discuss how the extracted rules can be used in both machine and human composition.

Machine learning models are useful and attractive tools for the interactive computer musician, enabling a breadth of interfaces and instruments. With current consumer hardware it becomes possible to run advanced machine learning algorithms in demanding performance situations, yet expertise remains a prominent entry barrier for most would-be users. Currently available implementations predominantly employ supervised machine learning techniques, while the adaptive, self-organizing capabilities of unsupervised models are not generally available. We present a free, new toolbox of unsupervised machine learning algorithms implemented in Max 5 to support real-time interactive music and video, aimed at the non-expert computer artist.

Supervised machine learning enables complex many-to-manymappings and control schemes needed in interactive performance systems. One of the persistent problems in theseapplications is generating, identifying and choosing inputoutput pairings for training. This poses problems of scope(limiting the realm of potential control inputs), effort (requiring significant pre-performance training time), and cognitive load (forcing the performer to learn and remember thecontrol areas). We discuss the creation and implementationof an automatic "supervisor," using unsupervised machinelearning algorithms to train a supervised neural networkon the fly. This hierarchical arrangement enables networktraining in real time based on the musical or gestural control inputs employed in a performance, aiming at freeing theperformer to operate in a creative, intuitive realm, makingthe machine control transparent and automatic. Three implementations of this self supervised model driven by iPod,iPad, and acoustic violin are described.

Description

We present an acoustic musical instrument played through a resonance model of another sound. The resonance model is controlled in real time as part of the composite instrument. Our implementation uses an electric violin, whose spatial position modifies filter parameters of the resonance model. Simplicial interpolation defines the mapping from spatial position to filter parameters. With some effort, pitch tracking can also control the filter parameters. The individual technologies -- motion tracking, pitch tracking, resonance models -- are easily adapted to other instruments., Proceedings of the International Conference on New Interfaces for Musical Expression, 2001