Latent Jamming is an improvisation practice with neural audio synthesizers that adopts concepts from algorithmic composition and generative music.
The following text is an excerpt from the companion paper to my contribution to the 7th Conference on AI Music Creativity (AIMC).
Training custom neural audio models
Latent Jamming positions itself in context with artistic practices in the “AI-Music” paradigm that explicitly include dataset curation and custom model training as part of a creative decision-making process (Jourdan et al., 2026; Tahiroğlu et al., 2024).
Preselection and categorization of music and audio data is considered a first creative act in the process, where material with a particular sonic character (e.g. sparse, dense, energetic, calm), attributed to a particular music genre, or such from a particular working phase or dedicated output selection (e.g. an album), is first separated into various data subsets and augmented in preparation for training.
Using architectures RAVE (Caillon et al., 2021), vschaos2 (Chemla—Romeu-Santos, 2020), or AFTER (Demerlé et al., 2024), among others, I train (variational) autoencoders on these curated selections with the aim of reproducing certain characteristics but even more importantly generating sounds that have not been part of the original dataset.
Observing models in action, the verdict is often that behavior and output can vary significantly, even when initial training conditions are similar or identical. This has contributed to a growing understanding of models possessing their own creative agency, while humans employing them in their work must share theirs (Xambó et al., 2024; Vear et al., 2023). Therefore, the models can be considered hybrids of instruments that embody the artist’s own aesthetics and sound-machines that partly act autonomously.
Latent space as entry point
In a composition-performance continuum, Latent Jamming follows an approach that embraces this notion of both model-as-instrument and (semi-)autonomous actor. Human-machine interaction and mutual influence are established and negotiated inside the models’ latent space, where low-level representations of the original domain data are accessible as latent embeddings or “encodings” for creative practitioners to engage with (Tahiroğlu et al., 2024; Yee-King, 2022)
In the past years, practitioners have found various ways of acting from within latent space, all with the intent of establishing a control layer (often including an interface) that aims for reproducibility and meaningful action-reaction patterns with regards to the models’ output. Usually, these techniques are titled in association with the terminology of space, e.g. “navigation” (Tahiroğlu et al., 2024), “mapping” (Zheng et al., 2026), or “exploration” (e.g. Liu et al., 2026; Tatar et al., 2023). Other approaches in latent space interaction make use of solutions like dimensionality reduction (e.g. Horta et al., 2025; Tahiroğlu et al., 2021) or aligning encodings of non-audio-domain control data along with latent embeddings of audio data (Tahiroğlu et al., 2026; Zheng et al., 2024).
Since in practice, each model requires statistical or empirical observation post-training as stated above, any composition-performance setup and its control interface can at most be a boilerplate template for techniques that have generally proven useful in similar cases, while putting it into action often resembles learning an instrument from scratch.
Learning an instrument in real-time
Latent Jamming builds on this instrument metaphor by using autoencoders’ decoder as a neural audio synthesizer. “Tuning” the instrument happens as an iterative process of generating arbitrary synthetic latent data, observing the decoded outcome in the audio domain followed by changing or stabilizing the initial input until an aesthetically convincing, yet more or less unstructured result has surfaced.
While this first step mainly works as initialization and exploration of the models’ general capabilities, a second step aims to structure the output by applying action patterns, modulation sequences, or simple repetition to synthetic data. In this context, the composition-performance practice becomes relevant, and the influence of human and artificial agents in a co-creative process is negotiated.
For Latent Jamming, I use Pure Data (PD) where real-time neural audio synthesis is made possible through nn~. This object exposes the number of latent dimensions a trained model has been exported with and allows observing and modulating latent trajectories created through its encoder, or injecting similar information into its decoder.
In PD, I programmed a set of custom abstractions that are tailored for nn~ compatible model types, serving as both a control interface for “playing” neural audio synthesizers as instruments and define the range within which the models can generate sound autonomously.
These abstractions can spawn and control signal streams that mimic latent embeddings and their trajectories while allowing to preserve constellations that have led to convincing results in the music-making process. Technically, they operate within a spectrum of random sampling from pre-defined distribution types to manual selection of data points both within a specified range.
Use cases and examples
In the past years, I’ve developed various composition-performance setups in Pure Data that build on the idea of Latent Jamming and allow exploring new ways of music co-creation, examples being Reykjavík Sunburn or Saatgut Proxy.
With UnRAVEl, I’ve started to transpose observations from Latent Jamming into a speculative composition practice.
References
Caillon, A., and Esling, P. (2021). RAVE: A variational autoencoder for fast and high-quality neural audio synthesis. 10.48550/arXiv.2111.05011.
Chemla—Romeu-Santos, A. (2020). Manifold representations of musical signals and generative spaces.
Demerlé, N., Esling, P., Doras, G., and Genova, D. (2024). Combining audio control and style transfer using latent diffusion. ArXiv, abs/2408.00196.
Horta Valenzuela, M., and Tomás, E. (2025). NEBULA: A PCA-Based Method to Explore RAVE-Encoded Audio Representations. In Proceedings of the 22nd Sound and Music Computing Conference (SMC 2025), Graz, 49–56.
Jourdan, T., Françoise, J., and Bevilacqua, F. (2026). Collective Craft: How Artists Collaborate to Train AI-based Audio Synthesis Model for Music. In Proceedings of the 2026 Conference on Creativity and Cognition (C&C 2026), New York, 921–933.
Liu, J., and Zheng, S. J. (2026). Exploring AI Audio Models in Soundwalking with Broader Audiences. In Proceedings of the 2026 Conference on Creativity and Cognition (C&C 2026), New York, 1699–1703.
Tahiroğlu, K., Kastemaa, M., and Koli, O. (2021). GANSpaceSynth: A Hybrid Generative Adversarial Network Architecture for Organising the Latent Space using a Dimensionality Reduction for Real-Time Audio Synthesis. In Proceedings of the 2nd Conference on AI Music Creativity (AIMC 2021).
Tahiroğlu, K., and Wyse, L. (2024). Latent Spaces as Platforms for Sonic Creativity. In Proceedings of the 15th International Conference on Computational Creativity (ICCC 2024), Sweden.
Tahiroğlu, K., Hokkanen, M., and Marta, A. (2026). Human-in-the-Loop: Crossmodal AI Alignment between Movement and Audio Latent Spaces for Expressive Sonification in Dance Performance. In Proceedings of the International Conference on New Interfaces for Musical Expression (NIME 2026), London.
Tatar, K., Cotton, K., and Bisig, D. (2023). Sound Design Strategies for Latent Audio Space Explorations Using Deep Learning Architectures. 10.48550/arXiv.2305.15571.
Vear, C., Benford, S., Avila, J. M., and Moroz, S. (2023). Human-AI Musicking: A Framework for Designing AI for Music Co-creativity. In Proceedings of the 4th AI Music Creativity Conference (AIMC 2023), Sussex.
Xambó, A. and Roma, G. (2024). Human–machine agencies in live coding for music performance. Journal of New Music Research, 53, 33–46.
Yee-King, M. (2022). Latent Spaces: A Creative Approach. In The Language of Creative AI: Practices, Aesthetics and Structures. Springer Series on Cultural Computing. Cham, 137–154.
Zheng, S., Sedó, A., and Bryan-Kinns, N. (2024). A Mapping Strategy for Interacting with Latent Audio Synthesis Using Artistic Materials. ArXiv, abs/2407.04379.
Zheng, S. J., Yoshida, K., García-Peguinho, N., Liu, J., Hearn, D., Xambó Sedó, A., and Bryan-Kinns, N. (2026). Latent Terrain: Adapting Neural Audio Autoencoders as Design Materials in NIME. In Proceedings of the International Conference on New Interfaces for Musical Expression (NIME 2026), London.