Koray Tahiroğlu, Miranda Kastemaa, Oskar Koli

    Research output: Artistic and non-textual formSoftwareScientificpeer-review


    GANSpaceSynth is a hybrid architecture for audio synthesis with deep neural networks that applies the GANSpace method to the GANSynth model. Using dimensionality reduction (PCA), we organise the latent space of trained GANSynth models, obtaining controls for exploring the range of sounds that can be synthesised.

    To perform GANSpace on GANSynth, we feed a large number of random latent vectors to the model and record the activations on an early convolutional layer of the network. We then use incremental PCA to compute the most significant directions in the activation space, as well as the global mean and standard deviation along each direction. When synthesizing, we give a coefficient for each direction, denoting how far to move along that direction (scaled by the standard deviation) starting from the global mean.
    Original languageEnglish
    PublisherAalto University; Aalto-yliopisto
    Media of outputOnline
    Publication statusPublished - 15 Jun 2021
    MoE publication typeI2 ICT software


    • Artificial Intelligence (AI)
    • Deep Learning
    • GAN
    • GANSpaceSynth
    • Deep Learning with Audio

    Field of art

    • Performance
    • Composition


    Dive into the research topics of 'GANSpaceSynth'. Together they form a unique fingerprint.

    Cite this