Generation and neural instruments

Mora includes optional paths that generate audio from a prompt, propose score material in the editor, and synthesize sound with neural models. None of them is required to write a pattern or to hear it.

The core pattern language does not require model downloads, a cloud account, an audio input, or Studio. Patterns, sound output, and many visuals do not need an ML model; ML is an optional set of tools for discovery, listening, generation, and ranking. Each tool on this page has its own environment, models, and licenses, and none is needed at startup.

conjure generates samples from a text prompt through a configured local generation sidecar. The current sidecar uses the Stable Audio 3 Small-Music model.

Generated audio needs to enter the configured sound library before it can be played by name. On the SuperCollider side, the SuperDirtMora Conjure component enables a sample-load OSC responder; enable it only in a trusted environment (see SuperDirtMora.components). See Samples and slicing for working with sample libraries.

The optional performer assistant uses the exported Mora vocabulary and typed form validation to propose score material in the editor. It requires a configured local model server.

rave and the nn.ar path synthesize sound with neural models. They need a valid model, plugin support, and explicit loading and preparation.

An advanced audio constructor is ready to play only once the companion advertises its capabilities and preparation succeeds; the Lisp object existing on its own is not enough.

Model weights are separate prerequisites with their own terms, and not all of them permit unrestricted commercial use. Mora’s MIT source license does not cover them. See Licensing and assets.

• Machine listening returns rendered audio to analysis for measured visuals and tools.
• Learned search uses MuQ and MuQ-MuLan embeddings for similarity and natural-language queries.
• PHALAR stem pairing ranks library candidates for compatibility with an anchor recording or a captured audible mix.

Learned search and PHALAR both choose among recordings that already exist; neither generates audio.