Browsing by Author "Huhtala, Ville"

Filter results by typing the first few letters
Now showing 1 - 3 of 3
  • Black-box modeling audio effects with Koopman-based neural networks
    (2024-09-28) Huhtala, Ville
     School of Electrical Engineering | Master's thesis
    In the last few years, neural network-based black-box modeling of non-linear audio effects has improved significantly. Present recurrent and convolutional neural models can model audio effects with long-term dynamics, but they require many parameters, thus increasing the processing time. This thesis presents a Koopman-linearised audio neural network structure that lifts a one-dimensional signal (mono audio) into a high-dimensional approximately linear state-space representation with non-linear mapping, and then uses differentiable biquad filters to predict linearly within the lifted state-space. Two models based on the general structure are compared against state-of-the art recurrent and convolutional models. Results show that the proposed models match the high performance of the state-of-the-art neural models while having a more compact structure, reducing the number of parameters by tenfold, and having interpretable components.
  • Digitaaliset viivelinjapohjaiset jälkikaikualgoritmit
    (2021-05-16) Huhtala, Ville
     Sähkötekniikan korkeakoulu | Bachelor's thesis
  • KLANN: Linearising Long-Term Dynamics in Nonlinear Audio Effects Using Koopman Networks
    (2024-04-16) Huhtala, Ville; Juvela, Lauri; Schlecht, Sebastian J.
     A1 Alkuperäisartikkeli tieteellisessä aikakauslehdessä
    In recent years, neural network-based black-box modeling of nonlinear audio effects has improved considerably. Present convolutional and recurrent models can model audio effects with long-term dynamics, but the models require many parameters, thus increasing the processing time. In this paper, we propose KLANN, a Koopman-Linearised Audio Neural Network structure that lifts a one-dimensional signal (mono audio) into a high-dimensional approximately linear state-space representation with nonlinear mapping, and then uses differentiable biquad filters to predict linearly within the lifted state-space. Results show that the proposed models match the high performance of the state-of-the-art neural models while having a more compact architecture, reducing the number of parameters by tenfold, and having interpretable components.