ReverbFX: A Dataset of Room Impulse Responses Derived from Reverb Effect Plugins for Singing Voice Dereverberation

Description:  
ReverbFX is a dataset of 1,846 room impulse responses (RIRs) designed to advance research in singing voice dereverberation, with a special focus on artificial reverberation used in music production. Unlike conventional RIR datasets recorded in real acoustic spaces or simulated via physical models, ReverbFX is derived exclusively from a diverse collection of professional reverb audio effect plugins, including Protoverb, SkyNet, TAL-Reverb-4, and Valhalla Supermassive.

Each RIR was generated by processing Dirac impulses through these plugins with factory and custom presets (randomized for further diversity). The resulting RIRs cover a wide range of decay times (RT60 spanning 0.31 to 52.08 seconds) and reverberation characteristics, capturing the creative breadth of artificial reverbs encountered in contemporary music production.

A stringent validation protocol was applied to ensure data quality, checking for numerical integrity, DC offset, signal energy, plausible RT60 range, envelope decay, and early energy concentration.

Key Features:  
- 1,846 plugin-derived RIRs  
- Four widely used reverb plugins (see paper for details)  
- Wide range of reverberation times and characteristics  
- Fully open metadata and link to corresponding plugin settings  
- All audio files share the same specifications:
    - Duration: 00:10 (10 seconds)
    - Audio channels: Stereo
    - Sample rate: 48 kHz
    - Bits per sample: 32-bit float  
- Designed for benchmarking and developing dereverberation and audio enhancement models for creative scenarios

Data Structure:
The dataset is organized as follows:

ReverbFX
├── LICENCE.txt
├── meta
│   ├── test.csv
│   ├── train.csv
│   └── valid.csv
├── test
│   ├── Protoverb_100_03.wav
│   ├── Protoverb_105_00.wav
│       ... 
│   └── Valhalla_WideVocalSwell_10.wav
├── train
│   ├── Protoverb_100_01.wav
│   ├── Protoverb_100_02.wav
│       ...
│   └── Valhalla_WideVocalSwell_14.wav
└── valid
    ├── Protoverb_103_01.wav
    ├── Protoverb_106_04.wav
        ...
    └── Valhalla_WideVocalSwell_04.wav

- LICENCE.txt: License information
- meta/: Metadata and split definitions (CSV files for train/valid/test)
- train/, valid/, test/: Folders containing RIRs (WAV files), named by plugin and preset  
  - All WAV files: 10 seconds stereo, 48 kHz sample rate, 32-bit float

Use Cases:  
- Training and evaluation of machine learning models for dereverberation
- Benchmarking source separation and audio enhancement technologies in artificial reverb scenarios
- Research in music information retrieval and creative audio processing

License:  
Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).  
Non-commercial research and educational use only. For details, see LICENCE.txt.

Attribution:  
Please cite the original paper in any derivative works or publications:

Julius Richter, Till Svajda, Timo Gerkmann, "ReverbFX: A Dataset of Room Impulse Responses Derived from Reverb Effect Plugins for Singing Voice Dereverberation," ITG Conference on Speech Communication, Berlin, Germany, Sept. 2025.

Please acknowledge the original plugin developers.

Contact:  
For questions or collaboration inquiries, contact {julius.richter,timo.gerkmann}@uni-hamburg.de.

Link:  
Project website: https://sp-uhh.github.io/reverbfx