Deep Learning Techniques for Music Generation

Abstract

8 As Richard Feynman coined it: "What I cannot create, I do not understand." 9 Initially codified in 1950 by Alan Turing and named by him the "imitation game" [191], the "Turing test" is a test of the ability for a machine to exhibit intelligent behavior equivalent to (and more precisely, indistinguishable from) the behavior of a human. In his imaginary experimental setting, Turing proposed the test to be a natural language conversation between a human (the evaluator) and a hidden actor (another human or a machine). If the evaluator cannot reliably tell the machine from the human, the machine is said to have passed the test. 10 This is to avoid the bias (synthetic flavor) of a computer rendered generated music. 11 Improvisation is a form of real time composition. 12 With some precaution, as this division is not that strict. 13 The general objective of integrating sub-symbolic and symbolic levels into a complete AI system is among the "Holy Grails" of AI. Chapter 2 Method In our analysis, we consider five main dimensions to characterize different ways of applying deep learning techniques to generate musical content. This typology is aimed at helping the analysis of the various perspectives (and elements) leading to the design of different deep learning-based music generation systems 1. 2.1 Dimensions The five dimensions that we consider are as follows. 2.1.1 Objective The objective 2 consists in: • The musical nature of the content to be generated. Examples are a melody, a polyphony or an accompaniment; and • The destination and use of the content generated. Examples are a musical score to be performed by some human musician(s) or an audio file to be played. 2.1.2 Representation The representation is the nature and format of the information (data) used to train and to generate musical content. Examples are signal, transformed signal (e.g., a spectrum, via a Fourier transform), piano roll, MIDI or text. 2.1.3 Architecture The architecture is the nature of the assemblage of processing units (the artificial neurons) and their connexions.

Deep Learning for Audio Signal Processing

“…to morph between different instrument timbres. Briot et al [145] provide a more in-depth treatment of music generation with deep learning.…”

High-Level Control of Drum Track Generation Using Learned Patterns of Rhythmic Interaction

“…In addition to the VAE-based methods for control over music generation processes mentioned above, a number of other studies have applied deep learning methods to address the problem of music generation in general, as reviewed in [4]. Drum track generation has been tackled using recurrent architectures [5,6], Restricted Boltzmann Machines [7], and Generative Adversarial Networks (GANs) [8].…”

In a Silent Way

“…Recurrent Neural Networks (RNNs) have been the predominant neural network architecture for music generation tasks [6], including drum sequence generators [20,29]. Recent research has shown however that Temporal Convolutional Network (TCN) models [37] can perform just as well or better in the analysis of sequential data [1].…”