LibriPara

Overview

LibriPara is an audio dataset designed for research on speaker-aware paralinguistic event detection and unified audio event detection.

It is associated with the paper:

SA-UAED: Joint Frame-Level Detection of Audio Events, Speaker Activities, and Speaker-Attributed Paralinguistic Events
Interspeech 2026.

The dataset is intended to support joint modeling of:

• audio events,
• speaker activities,
• speaker-attributed paralinguistic events,
• frame-level temporal localization.

Each sample is accompanied by the corresponding annotations and conversation-level metadata.

---

Dataset Scale

Split	Duration
Train	500 hours
Validation	5 hours
Test	5 hours
Total	510 hours

---

Dataset Structure

The repository is organized approximately as follows:

Libripara/
├── train/
│   ├── audio/
│   │   ├── 00000/
│   │   ├── 00001/
│   │   ├── ...
│   │   └── 000xx/
│   ├── labels/
│   │   ├── 00000/
│   │   ├── 00001/
│   │   ├── ...
│   │   └── 000xx/
│   └── conversations/
│       ├── 00000/
│       ├── 00001/
│       ├── ...
│       └── 000xx/
├── val/
│   ├── audio/
│   ├── labels/
│   └── conversations/
└── test/
    ├── audio/
    ├── labels/
    └── conversations/

To avoid storing too many files in a single directory, large subsets are divided into multiple shard directories such as:

00000/
00001/
00002/
...

Files belonging to the same sample share the same file stem and are stored in the corresponding shard.

For example:

train/audio/00003/example_001.wav
train/labels/00003/example_001.*
train/conversations/00003/example_001.*

These files correspond to the same sample.

---

Tasks

LibriPara can be used for research on:

• Unified Audio Event Detection
• Sound Event Detection
• Speaker Activity Detection
• Speaker Diarization
• Speaker-Aware Paralinguistic Event Detection
• Non-Verbal Vocalization Detection
• Laughter and Cough Detection
• Frame-Level Audio Event Localization
• Multi-Task Speech and Audio Understanding

---

Download

Hugging Face CLI

Install the Hugging Face Hub client:

pip install -U huggingface_hub

Then download the full dataset:

hf download originalover/Libripara \
    --repo-type dataset \
    --local-dir ./Libripara

The dataset will be saved to:

./Libripara

Python

You can also download the dataset using huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="originalover/Libripara",
    repo_type="dataset",
    local_dir="./Libripara",
)

---

Reading the Dataset

Because the training data are stored in multiple shard directories, recursive file traversal is recommended.

from pathlib import Path

root = Path("./Libripara")
audio_root = root / "train" / "audio"

audio_files = sorted(audio_root.rglob("*.wav"))

print(f"Number of training audio files: {len(audio_files)}")

for wav_path in audio_files[:5]:
    print(wav_path)

To locate the corresponding label and conversation metadata:

from pathlib import Path

root = Path("./Libripara")

audio_root = root / "train" / "audio"
label_root = root / "train" / "labels"
conversation_root = root / "train" / "conversations"

wav_path = next(audio_root.rglob("*.wav"))

relative_path = wav_path.relative_to(audio_root)
shard = relative_path.parent
sample_id = wav_path.stem

label_candidates = list(
    (label_root / shard).glob(f"{sample_id}.*")
)

conversation_candidates = list(
    (conversation_root / shard).glob(f"{sample_id}.*")
)

print("Audio:", wav_path)
print("Label:", label_candidates)
print("Conversation:", conversation_candidates)

The exact file extensions of labels and metadata depend on the released dataset version.

---

Recommended Usage

When implementing a custom PyTorch Dataset, recursively search the audio directory instead of assuming that all audio files are directly stored under train/audio/.

from pathlib import Path

self.audio_files = sorted(
    Path("Libripara/train/audio").rglob("*.wav")
)

When locating the corresponding label and conversation metadata, preserve the same shard path.

---

Citation

If you use LibriPara in your research, please cite the following paper:

@inproceedings{lan26_interspeech,
  title     = {{SA-UAED: Joint Frame-Level Detection of Audio Events, Speaker Activities, and Speaker-Attributed Paralinguistic Events}},
  author    = {Zekun Lan and Wangyou Zhang and Yanmin Qian},
  year      = {2026},
  booktitle = {{Interspeech 2026}},
  pages     = {1406--1410},
  doi       = {10.21437/Interspeech.2026-2486},
  issn      = {2958-1796},
}

---

License

Please refer to the licenses and terms of use of the original source datasets and resources used to construct LibriPara.

Users are responsible for ensuring that their use of this dataset complies with the corresponding licenses and terms.

---

Repository

Hugging Face:

https://huggingface.co/datasets/originalover/Libripara

For questions regarding the dataset, annotations, or benchmark settings, please open an issue in the Hugging Face dataset repository.