ARTPARK-IISc/Vaani-Noise-Event-Dataset

Vaani Noise Event Timestamps

🚧 Dataset Status: Actively Being Built Data is being uploaded in batches. Current coverage is a subset of the final planned corpus (~167 hrs. train). Star/watch this repo to be notified of updates.

Dataset Summary

Vaani Noise Event Timestamps is a derived dataset from Project Vaani, a large-scale multilingual speech initiative by IISc Bangalore and ARTPARK that captures India's linguistic diversity across all districts.

This dataset provides noise event annotations with fine-grained timestamps for the subset audio recordings from the Vaani corpus. Each entry identifies background noise categories along with their precise start and end times within the audio clip — enabling research in:

• Noise-robust Automatic Speech Recognition (ASR)
• Sound Event Detection (SED)
• Audio quality analysis in real-world Indian environments
• Speech enhancement and noise suppression

---

Noise Categories

The dataset covers 7 top-level noise categories with fine-grained subcategory timestamps:

#	Category	Example Events
1	animal	Barking, mooing, bird chirps, insect noise, cat, hen, goat
2	vehicle_traffic	Horns, engines, motorbikes, sirens, train, general traffic
3	baby_child	Crying, babbling, yelling, playing, child laughter
4	singing_music	Background music, singing, instruments, prayer, devotional
5	phone_signal_alarm	Ringtones, beeps, alarms, sirens, bells, doorbells
6	appliance_machine	Fans, mixers, TVs, mics, typing, clocks, machinery
7	human_non_speech	Breathing, lip smacks, coughs, sneezes, snoring, throat clearing

---

Dataset Structure

Data Fields

Field	Type	Description
audio	Audio	Audio recording
imageFileName	string	File name of the image prompt shown during recording
state	string	Indian state where the recording was collected
district	string	District within the state
duration	float64	Duration of the audio clip in seconds
language	string	Language spoken by the participant
annotationQuality	string	Annotation quality: verified_timestamps / unverified_timestamps / no_timestamps
isTranscriptionAvailable	bool	Whether a manual transcript exists for this clip
transcript	string	Transcription with inline noise tags (e.g., <horn>...</horn>)
NoiseCategory	list[string]	Top-level noise categories present in the clip
NoiseSubCategoryTimeStamp	list[dict]	Noise events with fine-grained timestamps (see schema below)

annotationQuality Values

Value	Description
verified_timestamps	Noise events with clean timestamps — start/end boundaries have mutual agreement between multiple annotators
unverified_timestamps	Noise events with timestamps that are annotated but inter-annotator agreement has not been verified
no_timestamps	Noise events are present but no timestamp boundaries are available; only the noise category is labeled

NoiseSubCategoryTimeStamp Schema

Each item in the list is a dictionary:

Key	Type	Description
category	string	Top-level noise category (e.g., vehicle_traffic)
tag	string	Inline transcript tag marking the noise (e.g., <horn>)
start	string	Start time of the noise event in seconds (e.g. "3.426")
end	string	End time of the noise event in seconds (e.g. "5.012")

Example Record

{
  "imageFileName": "Images/IISc_VaaniProject_GENERIC_0760.jpg",
  "state": "ArunachalPradesh",
  "district": "Longding",
  "duration": 1.418,
  "language": "Hindi",
  "annotationQuality": "verified_timestamps",
  "isTranscriptionAvailable": true,
  "transcript": "<horn> जो कि बच्चा लोग के लिए बनाया हुआ है। </horn>",
  "NoiseCategory": ["vehicle_traffic"],
  "NoiseSubCategoryTimeStamp": [
    {
      "category": "vehicle_traffic",
      "tag": "<horn>",
      "start": "0.013",
      "end": "1.399"
    }
  ]
}

---

Supported Tasks

Task	Description
Sound Event Detection	Detect and localize noise events within audio clips
Audio Classification	Classify the type of ambient noise in a recording
Noise-Robust ASR	Train/evaluate ASR models under real-world noise conditions
Speech Enhancement	Use timestamps to guide noise suppression systems

---

Usage

from datasets import load_dataset

ds = load_dataset("ARTPARK-IISc/Vaani-Noise-Event-Dataset")
print(ds)

# Access a sample
sample = ds["train"][0]
print(sample["NoiseCategory"])
# ['vehicle_traffic']

print(sample["NoiseSubCategoryTimeStamp"])
# [{'category': 'vehicle_traffic', 'tag': '<horn>', 'start': 0.013, 'end': 1.399}]

# Filter by noise category
vehicle_samples = [s for s in ds["train"] if "vehicle_traffic" in s["NoiseCategory"]]

---

Annotations

• Noise category labels and timestamps were annotated by trained annotators
• The transcript field contains inline noise tags (e.g., <horn>, <music>) marking where noise events occur within speech
• Timestamps are in seconds relative to the start of the audio clip
• A single clip can contain multiple overlapping noise events across different categories

---

Citation

If you use this data, please cite the following:

@misc{pulikodan2026vaanicapturinglanguagelandscape,
      title={VAANI: Capturing the language landscape for an inclusive digital India}, 
      author={Sujith Pulikodan and Abhayjeet Singh and Agneedh Basu and Nihar Desai and Pavan Kumar J and Pranav D Bhat and Raghu Dharmaraju and Ritika Gupta and Sathvik Udupa and Saurabh Kumar and Sumit Sharma and Vaibhav Vishwakarma and Visruth Sanka and Dinesh Tewari and Harsh Dhand and Amrita Kamat and Sukhwinder Singh and Shikhar Vashishth and Partha Talukdar and Raj Acharya and Prasanta Kumar Ghosh},
      year={2026},
      eprint={2603.28714},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2603.28714}, 
}

Contact Us

We are eager to hear your feedback about the dataset and are open to new ideas for collaborations as well. Feel free to reach out to us at vaanicontact@gmail.com