BanglaNoise-10: A Diverse Audio Dataset for Machine Learning on Urban Environmental Noise in Bangladesh

Authors: Md. Nasir Uddin, Md. Mehedi Hasan, and Mohammad Shahidur Rahman
Affiliation: Department of CSE, Shahjalal University of Science and Technology, Bangladesh

Abstract

Urban environmental noise data are critical for developing and evaluating audio analysis systems, yet publicly available datasets reflecting real-world urban noise in Bangladesh are limited. This data article presents BanglaNoise-10, an environmental audio dataset supporting research on urban noise analysis and sound processing.

Dataset Specifications

Specification	Details
Total Recordings	5,035
Duration per clip	10 seconds
Format	WAV (16 kHz, mono)
Categories	10 urban noise classes
License	CC BY 4.0

Dataset Classes

The dataset covers ten urban noise categories commonly observed in Bangladeshi cities:

1. Bike - Motorcycles, engine noise, acceleration, horn usage
2. Bus - Engine sounds, barking, acceleration, terminal noise
3. Car - Passenger vehicles, engine operation, horn usage
4. CNG Auto-rickshaw - Distinctive South Asian transport sounds
5. Construction Noise - Building sites, machinery
6. Protest - Culturally specific protest acoustics
7. Siren - Emergency vehicle sounds
8. Traffic Jam - Multi-layered dense traffic
9. Train - Railway sounds
10. Truck - Heavy vehicle noise

Data Collection

Recording Devices

• Samsung Galaxy S22 Ultra
• Oppo A15
• Realme C35

Collection Locations

Data collected across multiple regions of Bangladesh:

• Sylhet, Dhaka, Chattogram, Rajshahi
• Khulna, Barishal, Rangpur, Mymensingh
• Bandarban, Chandpur

Collection Period

February - November 2025 (10 months)

Machine Learning Baselines

Model Performance

Model	Accuracy	F1-Score
Whisper-Base	98%	High
Wav2Vec2-Base	97%	High
CNN	90%	Moderate

Training Configuration

Wav2Vec2-Base:

• Raw 16 kHz mono audio waveforms
• 10 epochs, AdamW optimizer
• Learning rate: $2 \times 10^{-5}$
• 80/20 stratified train-test split

Whisper-Base:

• Log-Mel spectrogram features
• 10 epochs, AdamW optimizer
• Learning rate: $3 \times 10^{-5}$

Value of the Data

1. First South Asian Dataset: First publicly available, large-scale environmental noise dataset for Bangladesh
2. Unique Regional Signatures: CNG auto-rickshaw, dense traffic jams, culturally specific protest acoustics
3. High Learnability: 98% accuracy demonstrates strong class separability
4. Raw Format: Unfiltered 16-bit WAV preserves full spectral content
5. Urban Computing Applications: Smart-city analytics, noise monitoring, health impact studies

Data Availability

• Repository: Zenodo
• DOI: 10.5281/zenodo.8239067
• URL: https://zenodo.org/record/8239067

CRediT Author Statement

• Md. Nasir Uddin: Data Collection, Preprocessing, Validation, Visualization
• Md. Mehedi Hasan: Conceptualization, Data Curation, Software, Visualization, Writing—Original Draft
• Mohammad Shahidur Rahman: Supervision, Project Administration, Resources

Keywords

Urban sound dataset, Audio dataset, Bangladesh, Sound classification, Machine learning, Acoustic analysis, Wav2Vec2, Whisper, Deep learning