Research map / Label-efficient and robust learning / Recognition and long-tail
DroneAudioset: An Audio Dataset for Drone-based Search and Rescue
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DroneAudioset, a 23.5-hour real-world multichannel drone audition dataset of human and ambient sounds recorded on a fixed-mounted drone at SNRs from -57.2 to -2.5 dB, and benchmarks MVDR/spectral gating, MPSENet, and SSLAM on it.
Problem
Human-presence detection from drone-mounted microphones is hampered by extreme ego-noise (SNR often below -10 dB), and existing drone audition datasets are small, low in SNR diversity, or synthetic.
Approach
A systematically controlled real-world recording setup (drone fixed to an aluminum frame, two 8-channel arrays plus a center mic, varied drones, throttles, distances, rooms, and source types) that yields a wide SNR range with separately recorded drone-only and source-only signals for SNR computation and augmentation.
Key results
- Neural (MPSENet) noise suppression gives the best SI-SDR for human vocal sounds at SNR above -10 dB (-12.37 dB vs -17.03 dB unprocessed).
- SSLAM classification F1 for human vocal sounds reaches 0.874 at SNR above -10 dB, approaching clean-recording performance of 0.968.
- All suppression methods perform poorly below -30 dB SNR, and non-vocal and ambient sound classes have low F1 (e.g., HNV 0.232, NH 0.192 at SNR above -10 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Neural enhancement rather than traditional beamforming and spectral gating | DroneAudioset noise suppression, HV sounds · SNR >-10 dB | SI-SDR (dB): -17.03 -> -12.37 | Helped |
| Use Neural enhancement rather than traditional beamforming and spectral gating | DroneAudioset noise suppression · -20 to -10 dB, HV/HNV/NH | SI-SDR (dB): -21.19 -> -14.30; -24.27 -> -16.33; -18.42 -> -15.80 | Helped |
| Use Hybrid rather than Neural enhancement | DroneAudioset noise suppression · SNR <-30 dB, HV/HNV/NH | SI-SDR (dB): -23.44 -> -22.76; -23.70 -> -22.77; -23.50 -> -22.85 | Helped |
| Use Hybrid rather than Neural enhancement | DroneAudioset classification, HNV · SNR >-10 dB | F1 score: 0.232 -> 0.280 | Helped |
| Use Hybrid rather than Neural enhancement | DroneAudioset classification, NH · SNR >-10 dB | F1 score: 0.192 -> 0.000 | Hurt |
| Increase input SNR | DroneAudioset human-presence classification · Across HV, HNV, and NH | F1 score | Helped |
| Place microphones below the drone rather than above it | DroneAudioset recording setup · Comparable throttle, source loudness, and microphone distance | Median SNR | Hurt |
| Increase drone-microphone distance from 25 cm to 50 cm | DroneAudioset microphone placement · High throttle; above-drone microphones | SI-SDR | Helped |
| Use the 8-channel array above the drone instead of the single-channel center microphone | DroneAudioset recording setup · High SNR, low throttle, 50 cm drone-to-microphone distance | SI-SDR | Helped |
| Reduce drone throttle from high to low | DroneAudioset recording setup, both drone sizes · Low-SNR setting | SI-SDR | Helped |
| Use the small drone rather than the large drone | DroneAudioset recording setup · Both throttle levels, including higher drone-source distances | SI-SDR | Helped |
Surprising findings
- Neural and hybrid suppression give little or no gain for non-vocal human and ambient sounds, with the hybrid method reaching F1 0.000 for NH at high SNR.
- Lower drone throttle and below-drone microphone placement can be viable under specific conditions, but the below-drone array has the lowest SI-SDR due to wind noise.
Limitations
- Drone is fixed to a frame rather than actually hovering, so flight micro-dynamics are not captured.
- Recordings are limited to three indoor rooms and a single azimuth; outdoor and cluttered conditions are not covered.
- Source sounds are played through a Bluetooth speaker, so the human sounds are not real human subjects.
Experimental setup
Datasets: AudioSet Freesound LibriSpeech Audio Distress Dataset DREGON (comparison)
Benchmarks: DroneAudioset
Metrics: SI-SDR F1 score
Baselines: MVDR beamforming with spectral gating MPSENet Hybrid (beamforming + MPSENet)
Builds on: MPSENet SSLAM MVDR beamforming SpeechBrain NoiseReduce
Closest papers by meaning
- AirCopBench: A Benchmark for Multi-drone Collaborative Embodied Perception and Reasoning (AAAI 2026 · similarity 0.82)
- MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking (NeurIPS 2025 · similarity 0.81)
- Massive Sound Embedding Benchmark (MSEB) (NeurIPS 2025 · similarity 0.81)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · similarity 0.80)
- EgoSound: Benchmarking Sound Understanding in Egocentric Videos (CVPR 2026 · similarity 0.80)
- Towards Audio-Visual Navigation in Noisy Environments: A Large-Scale Benchmark Dataset and an Architecture Considering Multiple Sound-Sources (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.