Generating datasets with audio plugins normally requires keeping track of exactly how each audio observation was created. A new arXiv paper introduces vst-bench-ml, a Python package designed for this purpose. The package samples plugin parameters and coordinates rendering through a separately installed host, which appears to decouple the sampling logic from the audio rendering environment.
The abstract emphasises that preserving the production process is a core requirement. By recording how each observation was generated, vst-bench-ml aims to support auditable and reproducible dataset creation for machine-learning benchmarks involving audio plugins. The paper is a single source, so there are no conflicting accounts to compare.
The package's approach of separating parameter sampling from host-based rendering may make it easier to trace a dataset back to its exact plugin settings and rendering conditions. This kind of provenance is useful for researchers who need to validate or reproduce experiments built on plugin-generated audio data.