Generating Albums with SampleRNN to Imitate Metal, Rock, and Punk Bands

Abstract

This early example of neural synthesis is a proof-ofconcept for how machine learning can drive new types of music software. Creating music can be as simple as specifying a set of music influences on which a model trains. We demonstrate a method for generating albums that imitate bands in experimental music genres previously unrealized by traditional synthesis techniques (e.g. additive, subtractive, FM, granular, concatenative). Raw audio is generated autoregressively in the timedomain using an unconditional SampleRNN. We create six albums this way. Artwork and song titles are also generated using materials from the original artists' back catalog as training data. We try a fully-automated method and a human-curated method. We discuss its potential for machine-assisted production.

Loopnet: Musical Loop Synthesis Conditioned on Intuitive Musical Parameters

“…Audio synthesis technologies for music have been researched for many years, ranging from synthesizers generating pitched waveforms, to singing voice synthesizers conditioned on melody and text, 3 to deep learning based models capable of generating entire songs [5,7]. For the context of this paper, we restrict ourselves to a description of generative models as deep learning based architectures used for musical audio synthesis.…”

“…The use of dilated causal convolutions allows the architecture to model longer term temporal dependencies between samples in an audio waveform than in the SampleRNN architecture. This architecture has subsequently been adapted for musical generation like singing voice synthesis conditioned on lyrics [11], and instrument sound generation conditioned on the pitch and latent representations of timbre [3,5,9]. While the output of these models is subjectively similar to natural-sounding samples, the sequential nature of the model means that the processing time for generation is quite high, unless high-resource processing units are available.…”

“…In parallel, recent advancements in generative deep learning based methodologies have sparked a huge interest in music generation. Research has focused on both the symbolic [1,2] and the audio domain [3][4][5][6][7][8][9]. However, except for singing voice synthesizers, where the parameters of synthesis can be clearly defined in terms This work is partially supported by the Towards Richer Online Music Public-domain Archives (TROMPA) project and partly by the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No765068, MIP-Frontiers.…”

“…of the lyrics and score [10][11][12], most of the proposed generative models for music rely on abstract latent embeddings for controlling the synthesis [4,5,[7][8][9]. Efforts have been made to find semantic structures within the latent embeddings that can provide an intuitive control to the music producer, but there remains a gap between the intuition behind the parameters of the generative model and the perceptual qualities of the audio generated.…”

Neural Network Layer Algebra: A Framework to Measure Capacity and Compression in Deep Learning

“…D EEP learning models have a good ability to deal with challenging problems that are too complex for us to explain by means of simple and deterministic laws in closed forms. Some examples include the extraction of relevant information from images [1], image inpainting and denoising [2], natural language processing [3], the creation of music [4], and learning how to play a 3D role-playing game properly [5].…”