Pregled bibliografske jedinice broj: 1152616
A Concept of a Wavetable Oscillator Based on a Neural Autoencoder
A Concept of a Wavetable Oscillator Based on a Neural Autoencoder // Audio Mostly (AM 2021)
online: The Association for Computing Machinery (ACM), 2021. str. 240-243 doi:10.1145/3478384.3478389 (predavanje, međunarodna recenzija, cjeloviti rad (in extenso), znanstveni)
CROSBI ID: 1152616 Za ispravke kontaktirajte CROSBI podršku putem web obrasca
Naslov
A Concept of a Wavetable Oscillator Based on a
Neural Autoencoder
Autori
Kreković, Gordan
Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni
ISBN
978-1-4503-8569-5
Skup
Audio Mostly (AM 2021)
Mjesto i datum
Online, 01.09.2021. - 03.09.2021
Vrsta sudjelovanja
Predavanje
Vrsta recenzije
Međunarodna recenzija
Ključne riječi
Sound synthesis ; wavetable synthesis ; deep learning ; autoencoder ; generative neural network
Sažetak
Inspired by the recently regained popularity of wavetable synthesis and rapid developments of neural networks, this study introduces a wavetable oscillator based on a standard autoencoder. The purpose of such a neural network is to generate diverse and novel single-cycle waveforms based on a small number of input parameters with sufficient computational efficiency. A consequence of using latent variables directly as input parameters is a smooth transition between the generated waveforms when changes of input parameters are small. To investigate the influence of datasets and hyperparameters on the output distribution, we conducted a set of experiments. The results suggest that even small and efficient generative models can successfully perform this task and produce an interestingly wide range of waveforms. Influence on possible shapes and sonic characteristics of the generated waveforms can be achieved using a specifically designed, synthetic dataset for model training.
Izvorni jezik
Engleski
