PaRIRset: Public Address Room Impulse Response Set

PaRIRset is a dataset of stereo room impulse responses (RIRs) measured in 40 professional concert venues, each equipped with a Public Address (PA) sound-reinforcement system (d&b audiotechnik, L-Acoustics, Nexo, TW Audio, Electro-Voice, and others). To our knowledge it is the first RIR dataset captured in professional concert venues through their PA systems, rather than in offices, classrooms, or lecture halls as in prior speech-enhancement RIR corpora.

PaRIRset was introduced in:

Enric Gusó, Xavier Serra. "CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation." ISMIR 2026.

Project page (paper, code, model weights, audio examples): https://enricguso.github.io/crowdioset_parirset

See also the companion dataset CrowdioSet, an audience-noise dataset (ambiences, events, and synthetic sing-alongs) built for the same Live Music Source Separation task.

Dataset contents

This repository contains the processed, augmented stereo RIRs used to train the models in the paper, downsampled to 44.1 kHz / 16-bit to match MUSDB18HQ / MOISESDB. The paper additionally describes raw mono and multichannel recordings at 48 kHz; see the project page for their availability.

parirset/
├── train/        2216 stereo WAV files (44.1 kHz), 32 PaRIRset venues + speech-enhancement RIRs
├── test/          160 stereo WAV files (44.1 kHz), 8 held-out PaRIRset venues
└── venues.csv     per-venue metadata (PA manufacturer, model, city, calibration, etc.)

Splits follow the paper: 8 venues were reserved and never used during training, forming the test split. Use the test split for both validation and evaluation — there is no separate validation split.

File naming convention

train/: {id}_{venueNum}_{venueName}_{tag}.wav

tag	meaning
d1_orig	Original beyerdynamic MM1 (flat-reference) RIR, unmodified ($D_1$)
d1_auN	$D_1$ polarity/gain-augmented variant (5 gains × direct/tail split × polarity flip)
d2_capNN	Raw signal from Zylia ZM-1 array capsule NN (of 19), each with its own small displacement ($D_2$)
d3_bfNN	Virtual microphone NN from the Zylia beamformer, random orientation/polar pattern ($D_3$)
GP_ACE, GP_MIT, GP_SLR28	Non-PaRIRset RIRs bundled in from the ACE Challenge, MIT IR Survey, and SLR28 speech-enhancement corpora, included for the "SE + PaRIRset" training configuration in the paper

test/: {id}_{venueNum}_{venueName}_test.wav — $D_1$-style augmentation only (20 variants × 8 held-out venues = 160 files).

venues.csv columns

id, name (matches venueName in filenames), type (indoors/outdoors), date, venue (venue proper name), city, address, website, manufacturer (PA manufacturer), main_model/main_qty, sub_model/sub_qty, frontfill_model/frontfill_qty, comments, beyer_calibration_110 (calibration value), opinion_score, beyer/zylia (whether that measurement device's recordings are included).

Measurement methodology

Each RIR was captured with a single exponential sine sweep ($f_1=20$ Hz, $f_2=20$ kHz, $T=6$ s) per PA side (left/right), measured at the Front of House (FOH) position, using a Zylia ZM-1 microphone array alongside a beyerdynamic MM1 measurement microphone. After deconvolution, RIRs were noise-gated (100 ms pre-onset, 50 ms fade-in/out) and the left/right channels were time-aligned via cross-correlation. The Zylia-derived RIRs were equalized against the beyerdynamic MM1 reference with a second-order high-shelf filter ($f_0=3.5$ kHz, gain $=-4$ dB, $Q=0.5$). Full details are in Section 3 of the paper.

Loading

from datasets import load_dataset

ds = load_dataset("enricguso/parirset")
# ds["train"], ds["test"]

Or directly with soundfile/torchaudio after cloning:

git clone https://huggingface.co/datasets/enricguso/parirset

License

Released under CC0 1.0 (public domain).

Citation

Paper: https://arxiv.org/pdf/2607.27828

@inproceedings{guso2026crowdioset,
  title={CrowdioSet and PaRIRset: Two Datasets Towards Live Music Source Separation},
  author={Gus{\'o}, Enric and Serra, Xavier},
  booktitle={Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR)},
  year={2026}
}

Acknowledgments

This work was financially supported by the Catalan Government through the ACCIO-Eurecat funding grant (Project TRAÇA: "IAGen" 2023-2026). We thank the band Cala Vento for allowing us to measure PaRIRset during their Brindis tour, and Pepe Ferrer from Global Audio Solutions for contributing a RIR measurement.