schism-audio/openslr-rirs

OpenSLR Room Impulse Responses & Noises

Quick Start

from huggingface_hub import snapshot_download

# Download the full dataset
path = snapshot_download(
    repo_id="schism-audio/openslr-rirs",
    repo_type="dataset",
)

Dataset Description

The OpenSLR Room Impulse Responses & Noises dataset (SLR28) is a collection of real and simulated room impulse responses (RIRs) along with point-source noise recordings. All audio is 16 kHz WAV. The dataset was created for training and evaluating far-field speech recognition and audio augmentation systems, and is equally valuable for adding realistic room acoustics to any dry audio signal.

The dataset includes:

• Real RIRs measured in 3 rooms of different sizes (small, medium, large)
• Simulated RIRs generated using image-method simulation across a variety of room geometries and source/receiver positions
• Point-source noises consisting of ambient noise recordings from real environments

This Hugging Face mirror contains 61,260 WAV files: 60,000 simulated RIRs, 417 real/isotropic RIR and noise files, and 843 point-source noise files.

This collection is widely used in speech and audio research for data augmentation via convolution-based reverb. For drum transcription and synthesis, convolving dry drum samples or stems with these RIRs produces realistic reverberant training data spanning a wide range of acoustic environments.

Dataset Structure

Directory Layout

RIRS_NOISES/
  real_rirs_isotropic_noises/    # Real measured RIRs + isotropic noises
  simulated_rirs/                # Simulated RIRs (various room geometries)
  pointsource_noises/            # Point-source ambient noise recordings

Data Fields

Field	Type	Description
audio	Audio	WAV file referenced by metadata.csv
collection	string	Top-level upstream collection
kind	string	Simulated RIR, real/isotropic RIR/noise, or point-source noise
room_size	string	Simulated room-size group when available
room_id	string	Simulated room identifier when available
subset	string	Upstream subset label when available

File Counts

Directory	Files
simulated_rirs/	60,000
pointsource_noises/	843
real_rirs_isotropic_noises/	417
Total	61,260

Usage Examples

Convolve a dry drum sample with a RIR

import numpy as np
import soundfile as sf
from scipy.signal import fftconvolve

# Load a dry drum hit and a RIR
dry, sr_dry = sf.read("dry_snare.wav")
rir, sr_rir = sf.read("RIRS_NOISES/real_rirs_isotropic_noises/rir_room1.wav")

# Resample if needed (RIRs are 16 kHz)
# Apply convolution reverb
wet = fftconvolve(dry, rir, mode="full")[:len(dry)]
wet = wet / np.max(np.abs(wet))  # Normalize

sf.write("wet_snare.wav", wet, sr_dry)

Batch augmentation with random RIRs

import os
import random
import numpy as np
import soundfile as sf
from scipy.signal import fftconvolve

rir_dir = "RIRS_NOISES/simulated_rirs"
rir_files = [os.path.join(rir_dir, f) for f in os.listdir(rir_dir) if f.endswith(".wav")]

def augment_with_reverb(audio, sr, rir_path):
    rir, sr_rir = sf.read(rir_path)
    wet = fftconvolve(audio, rir, mode="full")[:len(audio)]
    return wet / (np.max(np.abs(wet)) + 1e-8)

# Apply a random RIR to a dry signal
dry, sr = sf.read("dry_kick.wav")
rir_path = random.choice(rir_files)
wet = augment_with_reverb(dry, sr, rir_path)

Dataset Creation

Source Data

The dataset was compiled by Tom Ko et al. for the paper "A Study on Data Augmentation of Reverberant Speech for Robust Speech Recognition" (ICASSP 2017). Real RIRs were measured using swept-sine excitation in rooms of varying sizes at Aachen, Germany (Aachen Impulse Response Database and others). Simulated RIRs were generated using the image-source method with randomized room dimensions, absorption coefficients, and source/receiver placements.

Annotations

No annotations are provided. The dataset consists of raw audio files organized by type (real, simulated, noise).

Known Limitations

• Hub viewer limits: The dataset server can resolve the default split, but row/info preview endpoints may exceed the 10 MB response limit for this large AudioFolder mirror. Use snapshot_download or streaming for programmatic access.
• 16 kHz sample rate: All files are 16 kHz, which is standard for speech but lower than typical music production sample rates (44.1/48 kHz). Upsampling may introduce artifacts.
• Speech-oriented design: Room geometries and microphone placements were chosen for speech recognition research. They may not perfectly represent typical music recording or performance spaces.
• No RT60 metadata: Per-file reverberation time (RT60) values are not provided for all RIRs, making it harder to select RIRs by reverb characteristics without manual measurement.

Related Datasets

This dataset is part of the Drum Audio Datasets collection by schism-audio. Room impulse responses are useful for augmenting dry drum audio with realistic room acoustics. Related datasets:

• schism-audio/dechorate-rirs -- Real measured multi-channel RIRs with controlled acoustic conditions (dEchorate)
• schism-audio/e-gmd -- 444 hours of drum performances (dry) that benefit from RIR augmentation
• schism-audio/groove-midi-dataset -- Aligned MIDI + audio drum performances
• schism-audio/drum-percussion-kits -- One-shot drum samples suitable for convolution with RIRs

Citation

@inproceedings{ko2017study,
  title={A Study on Data Augmentation of Reverberant Speech for Robust Speech Recognition},
  author={Ko, Tom and Peddinti, Vijayaditya and Povey, Daniel and Seltzer, Michael L and Khudanpur, Sanjeev},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2017}
}

License

This dataset is released under the Apache License 2.0.

You are free to use, modify, and distribute this dataset for any purpose, including commercial use, subject to the terms of the Apache 2.0 license.