Broken Forecasts: Feedbacking Latent Generators for Sonic Instability

Broken Forecasts: Feedbacking Latent Generators for Sonic Instability Bła˙ zej Kotowski∗ Music Technology Group Universitat Pompeu Fabra Barcelona, Spain [email protected] Abstract Broken Forecasts is a live sound performance built around a custom generative system that interprets machine learning uncertainty as a design material. The system combines autoregressive prediction with a delayed feedback path, where past latents are transformed and fed back into generation, producing an unstable, performable trajectory that drifts from learned patterns. A neural synthesizer decodes these latent trajectories into audio in real time. This feedback disrupts learned rhythmic and structural patterns, giving rise to complex, glitch-inflected textures and unpredictable sonic evolutions. Rather than seeking coherence or control, the performance foregrounds recursive instability as a source of aesthetic possibility. By engaging with the feedback dynamics of autoregressive sequence prediction, Broken Forecasts proposes an alternative mode of interaction with generative models, treating them not as tools of prediction, but as uncertain, performable artifacts within a situated sonic practice. 1 Introduction Broken Forecasts is a live sound performance that treats Machine Learning (ML) as an unstable and performable material. Rather than seeking control or coherence, it explores feedback within a generative model as a mechanism uncovering its quasi-material 2 qualities. The system reroutes delayed and modulated latent codes back into its own input, creating a recursive loop that destabilizes generation. This controlled deviation pushes the system beyond its training distribution, aligning with the Active Divergence paradigm (Broad et al., 2021). This work treats machine learning not as a predictive tool, but as a malleable design material shaped by affordances such as stochasticity, instability, and opacity. It explores non-semantic forms of control, where the performer loosely nudges trajectories without specifying concrete symbolic intent. The use of a small-scale architecture and curated datasets responds to ethical and expressive concerns, resisting the extractivism of large AI systems. The performance asks: what kinds of listening and control emerge when we attune to generative models as recursive, misaligned processes rather than predictable tools? ∗https://blazejkotowski.com 2 The term suggests that algorithms and computational systems can be approached as if they possess material qualities that offer resistances, affordances, and expressive possibilities. Aspects such as interface behavior, parameter responsiveness, and data-induced dynamics are not strictly physical, yet function as material constraints that shape interaction and creative practice (Gerlek and Weydner-Volkmann, 2025). Proceedings of the 6th Conference on AI Music Creativity (AIMC 2025), Brussels, Belgium, September 10th-12th 2 Related Work This work aligns with artistic practices that subvert machine learning systems, framing failure, indeterminacy, and feedback as aesthetic strategies. The 2019 ISEA panel on “Machine Flaws in Generative Art” showed how glitches and breakdowns in AI models become material for expression (Boyé, 2019). Cascone’s framing of the post-digital aesthetic similarly positions glitch and failure as core to digital sound art (Cascone, 2000). Böhlen treats classification slippage as a site of discovery, where system errors become opportunities for critique and meaning-making. (Böhlen, 2021). Grba coins the term “tactical AI art,” where artists work with the assumptions and biases of AI systems to expose, challenge, or transform them (Grba, 2022). Rather than merely deploying systems for stylistic effect, such practices experiment with how models behave under pressure, in unfamiliar conditions, or when deliberately pushed into breakdown. Benjamin et al. conceptualize machine learning uncertainty as a defining material property of AI systems, introducing the notion of "thingly uncertainty", as a form of indeterminate relation between ML outputs and the world that lends itself to open-ended interpretation and aesthetic exploration (Benjamin et al., 2021). Musically, the piece draws on practices of temporal recursion. In "I Am Sitting in a (Latent) Room", a Variational Autoencoder (VAE) is used to recursively encode and decode audio in real time, echoing Alvin Lucier’s acoustic feedback piece (Shaheed and Wang, 2024). The system enables real-time manipulation of latent parameters, mirroring a practice commonly employed in musical feedback systems: processing the signal before it is fed back to the system. Together, these works position Broken Forecasts within a body of experimental AI investigation that embraces instability and treats machine learning not as a reproductive agent but as a malleable quasi-material artifact with its idiosyncratic qualities, which afford creative possibility. 3 System Architecture The system is built on the foundation of two ML models, as illustrated in Figure 1: 1. Neural Decoder: A decoder trained to synthesize audio from latent representations derived from training in tandem with its encoder counterpart, in a VAE setting (Kingma and Welling, 2019). The decoder features a DDSP architecture, but unlike typical DDSP implementations, which use pitch and loudness for control, this system conditions generation solely on the low-dimensional latent vector (Engel et al., 2020). As a result of VAE training, the latent space retains a smooth internal structure, which supports the interpolation of features. 2. Latent Sequence Model: The model draws on autoregressive sequence modeling approaches such as GPT-style transformers (Vaswani et al., 2017). An encoder-only transformer is trained on sequences of latent vectors to model temporal unfolding of the audio. The transformer predicts a fixed-length block of latent codes at a time, given a context buffer of previously generated codes. Crucially, tokenization is avoided, maintaining the latents continuous, allowing for smooth manipulation of the latent stream by the performer. In performance, the model operates as a closed generative loop. At each step, the Latent Sequence Model predicts a new latent vector block based on a context drawn from prior generations. However, rather than simply feeding its outputs back into the model unchanged—as in standard autoregressive systems—Broken Forecasts introduces a second temporal path. Latents from earlier in the context window are first selected at a delay offset ( τ ) and transformed by the Feedback Modulator, which applies operations such as offsetting, scaling, or inversion. These modulated latents are then added to the newly predicted ones, and the resulting composite is appended to the context buffer, introducing recursive instability into the generative process. Live performer input modulates both the feedback gain and delay parameters, steering the instability’s intensity and direction in real time. In parallel, the resulting latent vectors are passed through the Latent Codes Modulator before synthesis. This module offers the same operations—offset, scaling, inversion—but applies them transiently during decoding rather than feeding them back into the generative loop. These modulations shape 2 Figure 1: System architecture of Broken Forecasts. The system consists of the Temporal Dynamics Module (top) and the Synthesis Pathway (bottom). The Latent Sequence Model predicts new latent codes from a rolling Context Buffer. In parallel, earlier codes delayed by τ are processed through the Feedback Modulator, applying operations such as offset, scaling, or inversion. These modulated delayed codes are then combined with the current predictions and appended to the buffer, introducing recursive instability into the generation process. In the synthesis pathway, the Latent Codes Modulator applies similar transformations to the latent codes in real time, shaping the decoded audio without affecting the generative loop. Performer input controls both modulation stages, enabling dynamic influence over both structural evolution and timbral expression. the sound’s timbre without affecting temporal structure or system’s memory. The modulated latents are then decoded by the Neural Decoder, producing audio shaped by both system instability and performer input. 4 Artistic Outcomes In performance, the system supporting Broken Forecasts behaves like a shape-shifting instrument. At low feedback levels, the model produces coherent gestures reflective of the training data. Higher level of feedback introduces spectral anomalies, rhythmic irregularities, or timbral drift, depending on the feedback delay time and specific amount. The performer guides this process, occasionally resetting or seeding new inputs, but mostly shaping the unfolding instability by adjustment of feedback parameters and modulation of latent representations. Broken Forecasts proposes an approach to AI music performance centered on recursion, failure, and feedback. It treats machine learning models as unstable, quasi-material artifacts whose breakdowns are not errors, but compositional opportunities. By working with small datasets, embracing nonsemantic control, and looping predictions back into themselves, the piece opens new spaces for sonic exploration and interpretability. Rather than optimizing or mastering AI, this work engages it tactically, exposing its limits and listening to its deviations. In doing so, it contributes to an emergent field of critical, performative AI art that foregrounds its instability, uncertainty, and materiality. 3 References Benjamin, J. J., Berger, A., Merrill, N., and Pierce, J. (2021). Machine Learning Uncertainty as a Design Material: A Post-Phenomenological Inquiry. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, CHI ’21, pages 1–14, New York, NY, USA. Association for Computing Machinery. Böhlen, M. (2021). Classification, Slippage, Failure and Discovery. In Proceedings of the 9th Conference on Computation, Communication, Aesthetics & X, pages 15–29. Boyé, P. (2019). Machine Flaws in Generative Art. In Proceedings of the 25th International Symposium on Electronic Art, pages 713–716. Broad, T., Berns, S., Colton, S., and Grierson, M. (2021). Active divergence with generative deep learning - A survey and taxonomy. In de Silva Garza, A. G., Veale, T., Aguilar, W., and y Pérez, R. P., editors, Proceedings of the Twelfth International Conference on Computational Creativity, ICCC 2021, México City, México (Virtual), September 14-18, 2021, pages 227–236. Association for Computational Creativity (ACC). Cascone, K. (2000). The Aesthetics of Failure: "Post-Digital" Tendencies in Contemporary Computer Music. Computer Music Journal, 24(4):12–18. Engel, J. H., Hantrakul, L., Gu, C., and Roberts, A. (2020). DDSP: differentiable digital signal processing. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Gerlek, S. and Weydner-Volkmann, S. (2025). Materiality and machinic embodiment: A postphenomenological inquiry into ChatGPT’s active user interface. Journal of Human-Technology Relations, 3:1–15. Grba, D. (2022). Lures of engagement: An outlook on tactical ai art. In Proceedings of the 10th Conference on Computation, Communication, Aesthetics & X, pages 58–74. Kingma, D. P. and Welling, M. (2019). An introduction to variational autoencoders. Found. Trends Mach. Learn., 12(4):307–392. Shaheed, N. and Wang, G. (2024). I Am Sitting in a (Latent) Room. In Proceedings of the International Conference on New Interfaces for Musical Expression, pages 333–338. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R., editors, Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998–6008. 4