Skip to main content

Datasets in the Treble SDK

A dataset is a set of impulse responses (IRs) packaged with the metadata describing the acoustic conditions that produced each one — source and receiver geometry, simulation method, room properties, and IR length. Datasets are the entry point for working with IR data in the SDK: each is loaded as an IR Collection, the object on which all subsequent querying, filtering, and audio access is performed.

Why work with datasets​

The dataset abstraction separates the metadata (small, on the order of kilobytes) from the IR audio (large, held in the cloud and fetched on demand). This separation is the source of most of its utility:

  • Ready-made acoustic data. A dataset provides simulated IRs without requiring you to define geometry, assign materials, configure sources and receivers, or run simulations — useful as training and evaluation data, or as a baseline you extend with your own work.
  • Lightweight collaboration. Because the shareable artefact is a small metadata file rather than the audio, a dataset is an efficient way to pass acoustic data between teams. An engineer who has run a set of simulations can hand the file to a colleague in another discipline, who reloads it and retrieves the audio from the cloud on demand without the need for any large transfers or re-simulation.
  • A uniform interface. However a dataset is obtained, it behaves identically downstream, so purchased and internally built data can be combined and handled the same way.

Where datasets come from​

A dataset can be acquired in two ways:

  • Purchased from Treble. Curated datasets are available directly from Treble, spanning a range of room types, source–receiver configurations, and simulation methods. A free starter dataset, Treble10, is available for evaluation.
  • Built from your own simulations. Completed simulations can be assembled into a collection and persisted as a dataset in its own right — the basis for the cross-team sharing described above.

What an IR Collection lets you do​

Once loaded, a dataset is an IR Collection backed by a Polars DataFrame, with one IR per row. At a high level, this allows you to:

  • Query and filter IRs by their metadata (e.g. source–receiver distance, IR length, simulation method).
  • Enrich the collection with computed quantities and pre-computed acoustic parameters (T30, EDT, D50, C50, and others).
  • Shape a dataset through distribution-based sampling and preprocessing filter pipelines.
  • Retrieve audio — mono, spatial, moving, and device-rendered IRs — with automatic caching.
  • Feed downstream tasks, pairing each waveform with the acoustic conditions that produced it, which is the basis for audio AI workflows.

For complete method signatures and examples, see the working with IR collections how-to page.