PaperlayerView exact revision on arXiv

Source-linked AI summary

RIR-Mega: a large-scale simulated room impulse response dataset for machine learning and room acoustics modeling

Mandip Goswami

arXiv:2510.18917v2eess.ASeess.SP

TL;DR

Controlled room-acoustics studies can be difficult to support with data that is large-scale, well-described, and easy to reuse. RIR-Mega pairs simulated RIRs with a compact schema and tools for validation and access. A Random Forest baseline reports mean absolute error near 0.013 s and root mean square error near 0.022 s on its stated split.

  • Problem

    Controlled room-acoustics studies are hard to assemble at scale, and measured or simulated corpora can have limited metadata, unclear provenance, or inconvenient file layouts.

    Source: §1 · p. 1

  • Method

    RIR-Mega combines simulated shoebox-room RIRs with compact metadata, validation and reuse tools, and a reference RT60-like regression baseline.

    Source: §1 · p. 1, §3.1 · p. 2

  • Results

    On 36,000 training and 4,000 validation examples, a 400-tree Random Forest reaches mean absolute error near 0.013 s and root mean square error near 0.022 s.

    Source: §6 · p. 4

  • Takeaways & Limitations

    RIR-Mega combines simulated RIRs, a compact schema, validation tools, and a baseline as a practical resource for room-acoustics and robust-speech research.

    Source: §10 · p. 6

    All released examples are simulated shoebox rooms, which can differ from real spaces with non-ideal surfaces and complex geometries; the authors encourage testing on measured data when possible.

    Source: §9 · p. 5

Abstract

from arXiv · show

Room impulse responses are a core resource for dereverberation, robust speech recognition, source localization, and room acoustics estimation. We present RIR-Mega, a large collection of simulated RIRs described by a compact, machine friendly metadata schema and distributed with simple tools for validation and reuse. The dataset ships with a Hugging Face Datasets loader, scripts for metadata checks and checksums, and a reference regression baseline that predicts RT60 like targets from waveforms. On a train and validation split of 36,000 and 4,000 examples, a small Random Forest on lightweight time and spectral features reaches a mean absolute error near 0.013 s and a root mean square error near 0.022 s. We host a subset with 1,000 linear array RIRs and 3,000 circular array RIRs on Hugging Face for streaming and quick tests, and preserve the complete 50,000 RIR archive on Zenodo. The dataset and code are public to support reproducible studies.

1 Introduction

RIR-Mega addresses the difficulty of assembling controlled room-acoustics data by pairing simulated RIRs with a compact schema and practical reuse tools. Its contributions span data coverage, validation, a regression baseline, and distribution.

1 Introduction

  • Controlled studies are difficult to assemble at scale, while limited metadata and unclear provenance in existing corpora can leave researchers handling brittle scripts and file layouts.These burdens divert effort from the research questions themselves.Source: §1 · p. 1
  • A compact CSV or Parquet metadata file uses consistent paths and offers acoustic metrics as both a JSON-like string and flat columns, alongside a loader, validator, and checksum tool.The release also includes a baseline for RT60-like regression to help users sanity-check a pipeline.Source: §1 · p. 1
  • The overview traces a workflow from room specification through RIR generation and metadata to a loader and benchmarks.The Hugging Face subset and full Zenodo archive provide the stated distribution options.Source: §1 · p. 1, Figure 1 · p. 2
  • RIR-Mega contributes a large simulated RIR set with broad room-geometry and band-limited-decay coverage, plus a compact schema, validation tools, a regression baseline, and a two-part distribution plan.The distribution pairs fast Hugging Face streaming with long-term Zenodo archiving.Source: §1 · p. 1

2 Related work

RIR-Mega complements existing measured and simulated corpora with broad simulated coverage, a machine-readable schema, and modern distribution. The section specifies its simulation regime, room geometry, frequency-dependent absorption, and subset and archive sizes.

2 Related work

  • RIR-Mega complements existing corpora with large simulated coverage, a clear schema, and Hugging Face distribution, in a landscape where some resources have limitations in size, license clarity, or machine-readable metadata.A comparison table is deferred until scope and size are finalized across releases.Source: §2 · p. 2

3.1 Simulation regime

  • All released RIRs are simulated shoebox rooms with frequency-dependent absorption, generated using an image source method with configurable maximum reflection order.Each example records its sampling rate, reflection order, and metrics that can include RT60, DRR, clarity indices, and octave-band RT60s.Source: §3.1 · p. 2

3.2 Room geometry and placement

  • Room dimensions are sampled from Lx 2–25 m, Ly 3–30 m, and Lz 3–9 m, with sources and microphones placed inside the room away from walls.Microphone layouts are specified when the array field is set.Source: §3.2 · p. 2

3.3 Absorption and frequency bands

  • Frequency-dependent surface absorption is represented per band, with band-limited RT60 keys from 0.125 to 4.000 and scalar metrics including RT60, DRR, C50, and C80 when available.The listed octave-band keys are 0.125, 0.250, 0.500, 1.000, 2.000, and 4.000.Source: §3.3 · p. 2

3.4 Subsets and hosting

  • The Hugging Face preview subset contains 1,000 linear-array and 3,000 circular-array RIRs, while the canonical Zenodo archive preserves all 50,000 RIRs.The repository card explains how to switch between the subset and full archive.Source: §3.4 · p. 2

4 Metadata and schema

The metadata schema is compact and supports both CSV and Parquet. Common acoustic metrics are also exposed as flat columns for filtering, while the raw metrics string is retained.

4 Metadata and schema

  • The compact schema provides CSV and Parquet formats, with Parquet preferred for large tables and fast filters.The table caption identifies the schema as compact and notes flat numeric columns in Parquet.Source: §4 · p. 3, Table 1 · p. 3
  • Flat columns expose common metrics including RT60, DRR, C50, C80, and band RT60s, while preserving the raw metrics string.This offers direct access to common metrics alongside the original flexible representation.Source: §4 · p. 3

5 Loader and access

RIR-Mega provides a Hugging Face Datasets loader for a simple dataset interface. It resolves audio and prefers Parquet, with CSV as a fallback.

5 Loader and access

  • The `rirmega/dataset.py` loader exposes a simple interface and resolves audio through `datasets.Audio`.It prefers Parquet and falls back to CSV.Source: §5 · p. 3
  • The example loads RIR-Mega and retrieves a training example’s audio, path or array on access, and sample rate.It demonstrates the dataset interface in use.Source: §5 · p. 3

6 Benchmark

The benchmark provides a Random Forest baseline for RT60-like regression using lightweight waveform features, with errors reported on a 36,000-example training and 4,000-example validation split.

  • Features include simple energy statistics, a decay-slope proxy from the energy decay curve, and spectral centroid.Source: §6 · p. 3
  • Near 0.013 s mean absolute error and 0.022 s root mean square error are reported for a 400-tree Random Forest with seed zero.The split contains 36,000 training examples and 4,000 validation examples.Source: §6 · p. 4
  • The script accepts a specified target key or selects one from a fixed order beginning with rt60.Source: §6 · p. 4

7 Leaderboard and Community Submissions

The leaderboard uses pull requests to collect benchmark results in the dataset card, with run details and code links recorded for transparency and reproducibility.

  • A leaderboard table in the Hugging Face dataset card collects community benchmark results.Source: §7 · p. 4
  • Contributors submit results by pull request, including the command, seed, dataset tag, and a code link.Source: §7 · p. 4
  • The submission process is intended to keep results transparent, versioned, and reproducible for the community.Source: §7 · p. 4

8 Technical validation

Technical validation presents distributions and waveform-derived checks for the simulated RIRs, while external comparison with measured room responses remains planned.

8.1 Distribution sanity

  • Validation plots show RT60, direct-to-reverberant ratio, room volume, room-dimension distributions, an Lx-by-Ly scatter colored by Lz, and waveform and decay-curve examples.Source: §8.1 · p. 4, Figure 2 · p. 5

8.2 Self consistency of RT60

  • The self-consistency protocol compares metadata rt60 with waveform-derived estimates from Schroeder energy decay curves, using aligned windows and reporting correlation, mean absolute error, and root mean square error.Table 2 is a template whose italic values are to be replaced with results from a run.Source: §8.2 · p. 4, Table 2 · p. 4

8.3 Planned external validation

  • Measured RIRs are absent from this release; a future comparison with public measured datasets such as ACE Challenge or OpenAIR is planned to assess RT60 and DRR range similarity.Figure 3 currently illustrates the comparison format using linear- and circular-array simulated subsets.Source: §8.3 · p. 4

9 Usage notes

Usage guidance covers streaming subsets, reproducibility tools, and licensing, while noting that simulated shoebox rooms do not capture all real-space conditions.

  • The Hugging Face subset supports streaming for previews, while the full 50,000-RIR archive is available on Zenodo for complete experiments.Source: §9 · p. 5
  • A metadata validator and checksum script support file verification, and the dataset card documents the release tag.The baseline script creates a validation fold from training data if none is present.Source: §9 · p. 5
  • All audio and metadata are released under Creative Commons Attribution 4.0.Source: §9 · p. 5
  • All examples are simulated shoebox rooms, whereas real spaces can have non-ideal surfaces and complex geometries; the paper encourages testing on measured data when possible.Source: §9 · p. 5
  • Figure 3 compares RT60 and DRR distributions for linear- and circular-array subsets, with measured-RIR comparison planned for a later release.Source: Figure 3 · p. 6

10 Conclusion

RIR-Mega combines simulated RIRs, a compact schema, validation tools, and a working baseline as a practical resource for room acoustics and robust speech research. Its intended uses include dereverberation, room acoustics estimation, and robustness studies, while its scope remains limited by simulation and potential selection bias.

10 Conclusion

  • RIR-Mega combines simulated RIRs, a compact schema, validation tools, and a working baseline as a practical resource for room acoustics and robust speech research.The distribution plan balances fast access on the Hub with a permanent Zenodo archive, and the authors hope the resource reduces boilerplate work.Source: §10 · p. 6

11 Ethical Considerations, Usage, and Limitations

  • RIR-Mega is designed for dereverberation, room acoustics estimation, robustness studies, and related research.Source: §11 · p. 6
  • Selection biases may remain in room geometry distributions, material presets, and microphone arrays; specialized tasks may require parsing nested strings.The compact schema preserves free-form fields to maximize flexibility.Source: §11 · p. 6
  • Synthetic RIRs may not reflect all real-world complexities, so conclusions from purely simulated soundfields should be validated on measured data where feasible.Source: §11 · p. 6

Code availability

The repository provides the code and version information used for the manuscript, while the dataset card includes quick-start examples.

Loading 2510.18917v2…