BdNoise-10: Raw audio recordings for urban environmental noise classification

Overview BdNoise-10 is a large-scale environmental audio dataset containing 5,035 real-world urban noise recordings from Bangladesh. The dataset captures the chaotic and unique acoustic signature of developing urban environments, featuring 10 distinct categories: bike, bus, car, CNG auto-rickshaw, construction noise, protest, siren, traffic jam, train, and truck. Total Size: 5,035 files Format: WAV (16kHz, Mono, 16-bit) Duration: Fixed 10-second clips Geographic Coverage: 8 major cities in Bangladesh (Sylhet, Dhaka, Chattogram, Rajshahi, Khulna, Rangpur, Bandarban, Chandpur). Dataset Statistics Total Audio Files: 5,035 Total Duration: 839.17 minutes (~13.98 hours) Number of Categories: 10 File Format: WAV (16-bit, PCM) Sampling Rate: 16,000 Hz (16 kHz) Channels: Mono Data Sources: ~50% field recordings, ~50% online sources Category Distribution The dataset contains approximately 500-504 files per category for: bike, bus, car, cng_auto, construction, protest, siren, traffic_jam, train, and truck. File Naming Convention Audio files follow a strict naming pattern: {category}_{sequence_number}.wav Example: bike_0001.wav (First bike recording) Technical Specifications Audio Format: WAV (Waveform Audio File Format) Bit Depth: 16-bit signed PCM Sampling Rate: 16 kHz Recording Period: February - November 2025 Recording Devices: Samsung Galaxy S22 Ultra, Oppo A15, Samsung A30. Baseline Results Benchmarks established using an 80/20 train/test split: Whisper-Base: 99.1% Accuracy Wav2Vec2-Base: 97.3% Accuracy CNN: 90.2% Accuracy Contact For questions regarding this dataset: Md. Nasir Uddin ([e-mail]) Md Mehedi Hasan ([e-mail]) Mohammad Shahidur Rahman ([e-mail]) Department of CSE, Shahjalal University of Science and Technology (SUST), Bangladesh.