schismaudio/rirs-noises

RIRS NOISES

Quick Start

from datasets import load_dataset

# Stream to avoid downloading the entire dataset
ds = load_dataset("schismaudio/rirs-noises", streaming=True)

# Or download locally
ds = load_dataset("schismaudio/rirs-noises")

Dataset Description

RIRS NOISES is a collection of simulated and real room impulse responses (RIRs) plus isotropic and point-source noises from OpenSLR. The dataset aggregates RIRs from the RWCP Sound Scene database, the REVERB challenge, and the Aachen Impulse Response (AIR) database, alongside noise recordings from the MUSAN corpus. All audio is 16 kHz, 16-bit.

This dataset is designed for audio data augmentation — convolving clean audio with RIRs to simulate reverberant environments and adding noise at controlled SNRs. It is widely used for training robust speech and audio models.

Dataset Structure

Data Fields

Field	Type	Description
audio	Audio	WAV file at 16 kHz, 16-bit
filename	string	Original filename
type	string	Category: rir_simulated, rir_real, noise_isotropic, or noise_pointsource

Data Splits

This dataset has no predefined splits. All files are in the default train split.

Usage Examples

Load RIRs for augmentation

from datasets import load_dataset

ds = load_dataset("schismaudio/rirs-noises")

# Filter for real RIRs only
rirs = ds["train"].filter(lambda x: x["type"] == "rir_real")
print(f"Real RIRs: {len(rirs)}")

Convolve audio with an RIR

import numpy as np
from scipy.signal import fftconvolve
from datasets import load_dataset

ds = load_dataset("schismaudio/rirs-noises")
rir = ds["train"][0]["audio"]["array"]

# Normalize the RIR
rir = rir / np.max(np.abs(rir))

# Convolve with your dry audio signal
# reverberant = fftconvolve(dry_audio, rir, mode="full")

Dataset Creation

Source Data

The dataset was compiled from multiple publicly available acoustic databases:

• RWCP Sound Scene Database: Real-world RIRs recorded in various rooms and halls in Japan.
• REVERB Challenge: RIRs and multi-channel recordings for distant speech recognition research.
• Aachen Impulse Response (AIR) Database: Binaural and mono RIRs from diverse real environments.
• MUSAN: Music, speech, and noise recordings used for audio augmentation.
• Simulated RIRs: Generated using image-source methods for controlled room geometries.

Annotations

No manual annotations are included. File organization and naming encode the source database and recording type.

Known Limitations

• Speech-centric design: The RIRs and noise profiles were originally curated for speech processing. Drum and music applications may require additional filtering or selection.
• 16 kHz only: The fixed 16 kHz sample rate may be limiting for music applications that require higher fidelity.
• Simulated RIRs: The simulated subset uses simplified room models that may not fully capture real-world acoustic complexity.

Related Datasets

This dataset is part of the Drum Audio Datasets collection by schismaudio. Related datasets:

• schismaudio/dechorate — Calibrated multichannel RIRs with echo annotations and 3D positions

Citation

@misc{rirs_noises,
  title     = {A database of room impulse responses and noise recordings},
  author    = {Ko, Tom and Peddinti, Vijayaditya and Povey, Daniel and Seltzer, Michael L. and Khudanpur, Sanjeev},
  year      = {2017},
  url       = {https://openslr.org/28/}
}

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.