room-impulse-responses is an annotated index of acoustic RIR datasets

A list that describes the data, not just the link

room-impulse-responses is a list of publicly available acoustic and room impulse response datasets, and the README is explicit about the difference between it and the lists it is up against.

Unlike other (often outdated) lists, each entry briefly describes the dataset's content, so you don't have to download and inspect datasets one by one to find what you need.

That is the whole value proposition, and it is a smaller claim than it first appears. There is no data here. There is a table with a name, a description, a year, a paper or document, a licence and a flag for whether the download script covers it, and the descriptions are specific enough to save you the inspection work.

The scope is narrowed in two ways that are stated up front. Only real-world recorded RIRs are tracked here, so a synthetic or geometrically simulated response is out of scope by design. And some datasets also contain other data, for example speech or noise, but the descriptions cover their RIR content only, which is a useful piece of honesty since the extra material is what makes some of these archives large.

The index is maintained under the Graphi07 organisation, the primary language of the repository is Shell, the last push was on 2026-08-21, and the repository is not archived. There are no GitHub releases, so the index is a moving document rather than a versioned artefact.

What the measurements actually are

The entries are ordered newest first, and reading a few of them in order shows how wide the spread is.

FLAIR, from 2025, contains 270 room impulse responses in a single room, along with a millimeter accurate dense 3D point cloud of the room geometry and microphone positions, and it is positioned as a basis for investigating inverse spatial audio problems that rely on boundary information. That combination of recordings and geometry is unusual enough to be worth noting, because the geometry is what makes an inverse problem tractable at all.

SRIRACHA, also 2025, is much larger: 2,654,720 single-channel room impulse responses divided into eight different scenarios, from a combination of two source-receiver plane distances, two different source arrangements, and two different absorption levels of a shoebox room. It is intended as a benchmark for data-driven modelling and interpolation methods for room impulse responses.

MP-RIR, from 2024, has 68,736 responses measured on a dense 3D grid inside a complex-shaped room, and the measurement method is part of the entry: it was measured with a robot whose rotating arm operates as a linear guide and can carry a vertical linear array of eight omni-directional microphones, with four different sources at eight positions inside the room.

HOMULA-RIR, 2024, is for remote attendance teleconferencing, acquired with both higher-order microphones and a uniform linear array in a seminar room, where a 64-microphone ULA sat near two speakers while higher-order microphones modelled 25 attendees actually present in the room.

And then there are the older and differently shaped ones. MIRACLE, 2023, holds 856,128 single-channel responses across four measurement scenarios with a planar 64-channel array, including a regular grid of 64 by 64 source locations at two distances and a denser 33 by 33 grid at the short distance. SoundCam, 2023, covers 5,000 ten-channel measurements of room impulse responses plus 2,000 ten-channel music recordings in three rooms, a controlled lab, a living room and a conference room, with different humans positioned throughout each room. Arni, 2022, is 132,037 responses from 5,342 configurations of 55 acoustic panels. Motus, 2021, is 3,320 higher-order Ambisonic responses measured with an Eigenmike em32 across 4 loudspeaker positions and 830 furniture configurations, with spherical photographs and 3D models for each measurement. dEchorate, 2021, has 1,800 annotated responses from 6 arrays of 5 microphones, 6 sound sources and 11 acoustic conditions. MeshRIR, from the same year, is 4,410 mono responses recorded in very dense grids in a moderately reverberant space.

The variation in scale is the point: a project choosing a corpus is choosing between a few hundred responses with full geometry and a few million without it.

The download scripts, and the entries they skip

Download scripts are provided where possible, and datasets that require registration may not be linked directly. The entries in the table that can be fetched are marked, and the marked ones are downloaded by running the top level script with a destination folder:

./get_all_rirs.sh [destination folder]

The repository layout shows that there is one script per dataset alongside the combined one: get_but.sh, get_c4dm.sh, get_gtu_rir.sh, get_homula-rir.sh, get_miracle.sh, get_mird.sh, get_mit.sh, get_mprir.sh, get_openair.sh, get_rwcp_reverb_air.sh, get_soundcam.sh and get_sriracha.sh, next to get_all_rirs.sh.

That file list is more informative than it looks. Several of the names point at datasets that are not in the visible part of the table, so the scripts cover a wider set than the excerpt shows, and the naming follows the dataset rather than any convention, which means a script per dataset is the only sane structure for archives with different layouts.

The registration caveat explains the blank flag column. Some of these archives are on institutional repositories that require an account, and for those the index gives you the link and the description and expects you to fetch the data yourself.

For the ambisonic case there is a separate destination. For spatial RIRs in ambisonic format, that is the .sofa extension, the README points at a dedicated website at the SOFA conventions wiki. The project has drawn that line deliberately, since SOFA files carry spatial metadata in a structured container and are a different format from the plain per-channel recordings the scripts here handle.

The licence column is the real constraint

Every entry in the table carries a licence, and the spread across the column is wide enough to decide the project before you download anything.

MIT appears on SoundCam. CC BY 4.0 appears on FLAIR, MP-RIR, HOMULA-RIR, Arni, Motus and dEchorate. CC BY-SA 4.0 is on SRIRACHA. CC BY-NC-SA 4.0 is on MIRACLE, and GPL-3.0 is on GTU-RIR, which is described as more than 15,000 room impulse responses collected with a semi-automated sound recording system whose construction and extraction are explained in the linked GitHub repository.

The two entries to look at twice are the ones with non-commercial or copyleft terms. MIRACLE is CC BY-NC-SA 4.0, and NC means the data cannot be used for commercial purposes even though SA permits derivatives with the same licence. GTU-RIR is GPL-3.0, which for a dataset is unusual and has consequences if you redistribute it.

That is a practical trap in this field. A dataset of measured impulse responses is exactly the kind of asset that ends up inside a commercial product, a simulation engine or a training set, and the two most convenient corpora by size and by documentation happen to be the ones with restrictive terms. A research pipeline that only ever publishes results can treat most of the column as equivalent, and a product team cannot.

The BY-SA entries have their own requirement, since share-alike extends to adaptations, so a derived dataset inherits the obligation.

Where the index stops, and what it does not cover

There are three boundaries worth stating, because each one is a place where somebody will lose an afternoon.

The first is synthetic data. Only real-world recorded RIRs are tracked here, so a project that needs an analytically generated response, or a hybrid of simulated and measured responses, is outside the scope of this list by design rather than by omission.

The second is spatial audio in the SOFA format, which is explicitly delegated to the dedicated conventions website. That is a sensible division, since SOFA carries the spatial metadata in a structured way, but it does mean the index does not cover the ambisonic and higher-order spatial corpora as a single place, even though entries like Motus, described as higher-order Ambisonic responses measured with an Eigenmike em32, and HOMULA-RIR, acquired with higher-order microphones, are adjacent to that world.

The third is currency. The newest entry in the visible table is from 2025 and the list descends to 2021, and the repository has no GitHub releases, so there is no version to pin and no changelog to read. The list can therefore only be as current as its last commit, and the README's own comparison with other (often outdated) lists is a claim about descriptions rather than about freshness.

For a literature review or for choosing one corpus to start with, that is a good trade. For a project that intends to keep a long-lived dataset pipeline, a list with no release history and a table that changes in place is a dependency whose contents you have to re-check, and the licences have to be tracked per entry yourself, since the index does not encode them in a machine readable form.

Editorial conclusion

room-impulse-responses suits a researcher choosing a measured impulse response corpus for a simulation or a spatial audio experiment, since each row states the number of responses, the array geometry and the measurement conditions rather than only naming the dataset, and the scripts save a manual download pass over a dozen archives. It is the wrong tool if you need synthetic or simulated impulse responses rather than measured ones, because only real-world recorded RIRs are tracked here, and it is the wrong tool if you need spatial ambisonic data, which the project deliberately leaves to a separate SOFA resource. Before downloading, read the licence column for each dataset, because the entries range from MIT and CC BY 4.0 to CC BY-NC-SA 4.0 and GPL-3.0, and a non-commercial dataset cannot go into a commercial product.

Frequently asked questions

Official sources

1. Graphi07/room-impulse-responses on GitHub
2. Issues
3. README