Latent Terrain is a Max/MSP package by Jasper Shuoyang Zheng, developed over twelve months of performance practice and collaboration with four musician-researchers, and presented at NIME (New Interfaces for Musical Expression) in London in June 2026. The core operation: take a collection of your own sounds, run them through a neural audio codec (RAVE, Stable Audio, Music2Latent), record the latent vectors they produce, then train a small neural network to map spatial coordinates onto that latent territory. The result is a navigable 2D soundscape you can move through using a stylus, touchscreen, sensor, or controller — move slowly and timbres shift and drift; move quickly and they shatter, fracture, fuse into new textures. Four collaborators built actual works with the tool: Keigo Yoshida routed EEG brain signals into the terrain navigator while the sound space continuously retrained on incoming data, creating what he called "an adversarial tension between the performer's desire for calm and the system's pull toward arousal." Jiatong Liu built an ambient archive from field recordings of Beijing's rapidly disappearing Hutong neighborhoods and let listeners walk through it spatially. CDM featured the tool the same month.

The already-construct is the neural audio codec as compression infrastructure. RAVE, Stable Audio, Music2Latent — these models were designed to perform a specific task: compress audio into compact latent representations and reconstruct it. The entire AI audio generation ecosystem treats latent space as a means, not a destination. Input a prompt or a sound, pass through the latent space, arrive at output on the other side. Even the most experimental real-time transformation uses move through the codec's interior from A to B; nobody was meant to live there. Through training on vast audio data, these codecs built internal topologies — implicit maps of how sounds relate, cluster, and transition — as a byproduct of learning to compress efficiently. That topology was never intended as artistic territory. The chisel Latent Terrain performs is to suspend the codec's purpose entirely. It doesn't ask the codec to encode or reconstruct. It asks: what is the shape of the space the codec learned? Feed it your sounds, let the spatial geometry emerge, then hand that never-meant-to-be-traversed interior map to a performer. The remainder is the codec's understanding of sound — its learned timbral topology, distilled from training data — which was always there but was never the point.

Every one of the four collaborators describes working with unpredictability as a central fact of the experience, not a problem to be solved. "Active listening as a compositional approach — you work with what the space gives you." "Letting the system trace and access past states in real-time, not to establish full control over the outcome." "Learning to deal with the unpredictability" as a design question, not an obstacle. This is not a polite way of saying the tool is difficult to use. It is the direct phenomenological encounter with a remainder: a space whose organizing logic was distilled by machine learning from vast human-produced audio, whose structure is real and consistent, but was not designed for human traversal. The performer enters territory that belongs to no existing musical discipline. The rules are there — they were learned — but they were not learned for this.

The construct "AI audio = type what you want, receive output" is solidifying right now. As that cultural logic hardens, the imaginative space for inhabiting a codec's interior topology becomes harder to access — it starts to look like technique rather than the discovery of genuinely unnamed territory. NIME 2026 is the beginning of the academic naming process, but it has not finished: Max for Live devices are still inbound, support for additional codecs is expanding, a Pure Data version is being considered. The naming gap is acute. This is not AI generation (no prompts), not synthesis (no oscillators), not sampling (structural geometry rather than audio material), not FluCoMa-style corpus synthesis (FluCoMa maps sounds by audio analysis descriptors; Latent Terrain maps by learned latent topology — a different epistemology entirely). Zheng himself writes: "I'm not particularly interested in typing prompts to make stuff, I'm interested in breaking them and dissecting them, and when you crack them open, what's inside turns out to be genuinely playable." That is the remainder discoverer's sentence: go find the space no one has moved into yet, and make it an instrument.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics