DroneAudioset: An Audio Dataset for Drone-based Search and Rescue (2510.15383v1)
Abstract: Unmanned Aerial Vehicles (UAVs) or drones, are increasingly used in search and rescue missions to detect human presence. Existing systems primarily leverage vision-based methods which are prone to fail under low-visibility or occlusion. Drone-based audio perception offers promise but suffers from extreme ego-noise that masks sounds indicating human presence. Existing datasets are either limited in diversity or synthetic, lacking real acoustic interactions, and there are no standardized setups for drone audition. To this end, we present DroneAudioset (The dataset is publicly available at https://huggingface.co/datasets/ahlab-drone-project/DroneAudioSet/ under the MIT license), a comprehensive drone audition dataset featuring 23.5 hours of annotated recordings, covering a wide range of signal-to-noise ratios (SNRs) from -57.2 dB to -2.5 dB, across various drone types, throttles, microphone configurations as well as environments. The dataset enables development and systematic evaluation of noise suppression and classification methods for human-presence detection under challenging conditions, while also informing practical design considerations for drone audition systems, such as microphone placement trade-offs, and development of drone noise-aware audio processing. This dataset is an important step towards enabling design and deployment of drone-audition systems.
Sponsored by Paperpile, the PDF & BibTeX manager trusted by top AI labs.
Get 30 days freeSummary
- The paper introduces DroneAudioset, a dataset of 23.5 hours of audio recordings capturing varied SNR conditions for drone-based search and rescue.
- The methodology uses a two-stage processing pipeline combining noise suppression and human-presence detection with state-of-the-art metrics like SI-SDR.
- The findings provide design recommendations on optimal microphone placement and drone configurations to enhance audio classification performance.
DroneAudioset: An Audio Dataset for Drone-based Search and Rescue
Introduction
The paper presents "DroneAudioset", a comprehensive and systematic dataset designed to address the challenges of using audio data in search and rescue operations conducted by drones. The predominant reliance on visual data in such scenarios faces limitations under adverse visibility conditions. Auditory data offers a complementary approach, but the deployment of drone-mounted microphones is hindered by challenges like intense ego-noise generated by the drones themselves. Existing datasets are either limited in scope or synthetic, not adequately reflecting real-world conditions. DroneAudioset addresses these gaps by offering a rich dataset comprising 23.5 hours of audio recordings under varied conditions of signal-to-noise ratios (SNRs), drone configurations, and environments. This dataset is a crucial step toward developing robust drone audition systems capable of effective human presence detection in challenging scenarios.
Experimental Setup and Data Collection
DroneAudioset was meticulously designed and collected in controlled environments to simulate realistic drone usage scenarios in search and rescue operations. The dataset captures audio using unmanned aerial vehicles (UAVs) such as DJI F450 and DJI F330, which are mounted on a fixed frame to ensure consistent recording conditions (Figure 1).
Figure 1: (a) Experimental setup with the drone on an aluminum frame, (b) actual setup showing the drone frame, microphone array, and drones.
Two 8-channel microphone arrays and one standalone microphone were used to capture a wide range of audio, enabling recordings under varying conditions of drone throttle levels, microphone placement, and environmental setups. A total of 23.5 hours of audio data was collected, including drone noise, human sounds (vocal and non-vocal), and ambient sounds, across three different environments with varying reverberation characteristics.
Signal-to-Noise Ratio Analysis
Signal-to-noise ratio (SNR) is a critical metric in assessing the quality and usability of audio recordings in noisy drone environments. DroneAudioset captures a wide range of SNRs, from highly challenging conditions of -57.2 dB to more manageable levels of -2.5 dB (Figure 2). The dataset enables detailed analysis of how different drone configurations and environmental setups affect SNR, providing insights into designing effective noise suppression and audio classification models.
Figure 2: (a) Histogram of SNRs for all data, (b-g) SNRs across microphones and throttle levels.
Methodology for Noise Suppression and Classification
To process and utilize the captured audio effectively, the paper outlines a two-stage processing pipeline (Figure 3). This consists of noise suppression followed by human-presence detection through classification.
Figure 3: Pipeline for processing audio recordings, including noise suppression and human-presence detection.
The dataset was evaluated using various state-of-the-art noise suppression techniques, including traditional beamforming and neural enhancement methods. Performance was measured using the Scale-Invariant Signal-to-Distortion Ratio (SI-SDR), indicating the extent of improvement over the original recordings. Additionally, a classification system using SSLAM (Self-Supervised Learning from Audio Mixtures) was tested, mapping audio events to human vocal, non-vocal, or ambient sound categories.
Evaluation and Findings
The evaluation results demonstrate the efficacy of neural and hybrid methods for noise suppression, achieving notable SI-SDR improvements, particularly for human vocal sounds. However, more complex ambient sounds presented challenges, underscoring the need for further innovations in audio processing for drone environments. Classification performance varied significantly based on the noise suppression stage, highlighting the importance of continuous improvement across the entire processing pipeline.
Design Recommendations for Drone-Audition Systems
The dataset also informs key design considerations for future drone-audition systems:
- Microphone Placement: Above-drone configurations offer better noise suppression but below-drone positions can capture ground-level sounds more effectively.
- Drone Configuration: Larger drones with adaptive throttle management can mitigate noise impact during critical audio detection tasks.
- Microphone Configurations: Beamforming using multi-microphone arrays is beneficial but must be balanced with processing demands and payload capacity.
Conclusion
DroneAudioset represents a significant advancement in enabling effective drone audition systems for search and rescue operations. By providing a robust, open-source dataset, it allows researchers to test and develop advanced algorithms under realistic conditions. Future research will focus on addressing low-SNR challenges, integrating visual and auditory data, and expanding the dataset to outdoor environments. The development of drone-audition systems promises significant improvements in search and rescue capabilities, enhancing the ability to detect and respond to emergencies efficiently.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
What is this paper about?
This paper introduces DroneAudioset, a big collection of real audio recordings made with drones to help find people during search-and-rescue missions. Instead of only relying on cameras (which struggle in smoke, darkness, or clutter), the idea is to use microphones on drones to “hear” signs of people—like speech, screams, knocks, or footsteps. The catch: drones are extremely loud, and their own noise often drowns out important sounds. This dataset helps researchers build and test better methods for cleaning up that noise and detecting human presence.
What questions does the paper try to answer?
The authors focus on simple but important questions:
- Can we build a realistic, diverse audio dataset from drones that captures how hard it is to hear people around them?
- Which microphone positions and setups work best on a noisy drone?
- How well do today’s noise-cleaning and sound-classifying AI models work in these tough conditions?
- What practical tips can we offer for designing drone audio systems that actually work in the field?
How did they do it?
Think of a drone like a very noisy fan that also creates wind. That noise (called “ego-noise”) makes it hard to hear anything else.
To paper this properly, the team:
- Mounted two different drones on a sturdy frame to imitate hovering safely and consistently.
- Placed 17 microphones in different spots—above the drone, below the drone, and at different distances (25 cm and 50 cm)—including two round 8-microphone “arrays” (like having many ears) and one single microphone.
- Played different kinds of sounds from a speaker at various distances and volumes:
- Human vocal (speech, screams, crying)
- Human non-vocal (clapping, knocking)
- Non-human ambient (fire crackling, water dripping)
- Recorded:
- Drone + source together (most of the data)
- Drone noise alone
- Source sounds alone
- This lets them measure how loud the signal is compared to the noise, known as the Signal-to-Noise Ratio (SNR). Negative SNR means the noise is louder than the thing you want to hear. In this dataset, SNR often ranges from about −57 dB to −2.5 dB—very tough conditions.
They also tested algorithms to clean the audio (“noise suppression”) and then classify it to detect human presence:
- Traditional “beamforming” (like cupping multiple ears and pointing them toward the sound you care about) plus a method that mutes quieter background noise.
- A modern AI “neural” model that tries to separate speech from noise.
- A hybrid that first beamforms, then applies the neural cleaner.
- After cleaning the audio, an audio classifier (pretrained on many everyday sounds) tries to label the sound as human vocal, human non-vocal, or non-human.
Simple translations of technical terms:
- Ego-noise: the drone’s own noise from motors and propellers.
- Beamforming: using many microphones together to “focus” on sound coming from a certain direction—like a sound flashlight.
- Neural enhancement: an AI tool that learns patterns to remove noise and keep the important sound.
- SNR (Signal-to-Noise Ratio): how loud the thing you care about is compared to the background noise. Lower (especially negative) is worse.
- SI-SDR: a score for how close your cleaned-up audio is to a clean reference; higher is better.
- F1-score: a way to measure classification accuracy that balances “how many did you catch” and “how many did you get right.”
What did they find, and why does it matter?
Here are the key takeaways:
- The dataset is large and realistic:
- 23.5 hours of audio.
- Many setups: 2 drone sizes, 2 throttle levels (low/high), 17 microphones, various positions, 3 different rooms, multiple distances and sound loudness levels.
- SNRs are very low (often negative), matching real-life difficulty.
- Noise cleaning is hard under extreme drone noise:
- AI-based cleaning generally beats traditional methods, especially when the signal is very weak.
- Even so, when the noise is extremely strong (e.g., SNR below −30 dB), all methods struggle.
- Detecting human voice works better than detecting other sounds:
- Human vocal sounds (like speech or screams) are recognized more accurately after cleaning than non-vocal human sounds (like knocks or claps) or ambient sounds.
- At higher SNR (less noise), results get closer to clean-audio performance.
- Microphone placement matters:
- Mics above the drone usually work better than mics below it because wind from the propellers hits the lower mics directly.
- Increasing the mic’s distance from the drone (from 25 cm to 50 cm) often helps.
- Multi-microphone arrays can improve results (via beamforming), but they add weight and need more processing power.
- Flight choices matter:
- Lower throttle (less propeller power) means less noise and better audio.
- The larger drone tended to allow better results than the smaller one, likely due to different noise profiles and payload options.
Why this matters: Real search-and-rescue drones can’t always “see” people, but they might “hear” them—if we design the system right. This dataset gives researchers a standard, realistic way to test their ideas and improve drone hearing.
What’s the bigger impact?
- For researchers: This is a public, realistic dataset to build better noise suppression and sound detection methods that can handle extremely noisy drone audio. It also supports studying microphone placement and system design choices.
- For engineers and rescuers: The paper offers practical guidelines:
- Prefer microphones above the drone, and place them a bit farther from the propellers.
- Use multi-mic arrays when you can afford the extra weight and power.
- Reduce throttle during “listening moments” to cut noise.
- Consider drone size and payload limits when planning audio gear.
- For society: Better drone hearing could help save lives by finding people when cameras can’t. The authors also note privacy concerns and suggest ethical safeguards.
The authors also share limits and next steps:
- Their setup simulates hovering by mounting the drone, so future work should capture real flight dynamics.
- Most recordings are indoors; outdoor recordings (with stronger wind and bigger distances) are a valuable next step.
- More diverse environments and end-to-end systems tailored to non-vocal sounds would improve performance.
In short: DroneAudioset is a solid first step toward drones that can “listen” for people in dangerous places, helping rescuers act faster and more safely.
Collections
Sign up for free to add this paper to one or more collections.