Datasets:

mandipgoswami
/

LibriRIR-100

language:
  - en
license: cc-by-nc-4.0
tags:
  - automatic-speech-recognition
  - speech-enhancement
  - dereverberation
  - robustness
  - audio
  - reverberation
  - librispeech
  - room-acoustics
  - paired-data
  - training-data

LibriRIR-100

Dataset Summary

LibriRIR-100 is a large-scale paired clean↔reverberant speech training corpus containing exactly 100 hours of speech. Each utterance is paired with a room impulse response from RIR-Mega (mandipgoswami/rirmega), stratified across four RT60 reverberation conditions. Designed as a drop-in training resource for robust ASR, speech enhancement, and dereverberation models.

Why LibriRIR-100

Existing paired reverberant speech datasets are either small-scale, lack acoustic metadata, or are not freely available. LibriRIR-100 fills this gap by providing:

Exactly 100 hours of paired clean and reverberant speech
RT60/DRR/C50 metadata on every sample for condition-specific training and evaluation
Stratified RT60 bins (short, medium, long, very_long) for balanced acoustic diversity
Reproducible RIR assignment — each sample_id encodes the exact RIR used

Use Cases

Training robust ASR models — e.g. fine-tuning Whisper for reverberant conditions
Training dereverberation models — WPE, Demucs, neural dereverberation
Training speech enhancement models — denoising, dereverberation, super-resolution
Acoustic condition-aware model training — using RT60/DRR metadata for curriculum or multi-task learning

Dataset Structure

Column	Type	Description
sample_id	string	`{speaker_id}_{chapter_id}_{utterance_id}_{rir_id}`
audio_clean	Audio(16kHz)	Clean 16kHz mono FLAC
audio_reverb	Audio(16kHz)	Reverberant 16kHz mono FLAC
text	string	Ground truth transcript (lowercased)
speaker_id	string	LibriSpeech speaker ID
chapter_id	string	LibriSpeech chapter ID
utterance_id	string	LibriSpeech utterance ID
rir_id	string	RIR-Mega sample_id
rt60_bin	string	short / medium / long / very_long / unknown
RT60_T30_s	float	RIR RT60 (null if missing)
DRR_dB	float	RIR DRR (null if missing)
C50_dB	float	RIR C50 (null if missing)
duration_s	float	Utterance duration in seconds
split	string	train or validation

Splits: 90% train, 10% validation (stratified by rt60_bin).

RT60 bin distribution: Approximately 25% each for short, medium, long, very_long.

How It's Built

Speech source: LibriSpeech train-clean-100 (100h subset, CC BY 4.0)
RIR source: RIR-Mega v2 (mandipgoswami/rirmega)
Pipeline: Full reproduction code at github.com/mandip42/LibriRIR-100

Quickstart

from datasets import load_dataset

ds = load_dataset("mandipgoswami/LibriRIR-100")
sample = ds["train"][0]
# sample["audio_clean"], sample["audio_reverb"], sample["text"], sample["RT60_T30_s"]

Reproducing This Dataset

Clone the pipeline and run:

git clone https://github.com/mandip42/LibriRIR-100.git
cd LibriRIR-100
pip install -e .
python scripts/build_and_publish.py --config configs/default.yaml

Limitations

English only (LibriSpeech)
Single RIR per utterance (no multi-condition augmentation in this release)
RIR-Mega metadata (RT60, DRR, C50) may be missing for some samples

Citation

@misc{goswami2025libririr100,
  title        = {LibriRIR-100: A Paired Clean-Reverberant Speech Training Corpus},
  author       = {Goswami, Mandip},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/mandipgoswami/LibriRIR-100}
}

@misc{goswami2025rirmega,
  title        = {RIR-Mega: A Large-Scale Room Impulse Response Corpus},
  author       = {Goswami, Mandip},
  year         = {2025},
  eprint       = {2510.18917},
  archivePrefix= {arXiv}
}

License

CC BY-NC 4.0. Audio content derived from LibriSpeech (CC BY 4.0) and RIR-Mega (CC BY-NC 4.0).