DroneAudioset: An Audio Dataset for Drone-based Search and Rescue

TL;DR

DroneAudioset addresses the challenge of detecting human presence with audio in drone-based indoor search and rescue under extreme ego-noise. It introduces a real-world, systematically collected dataset of $23.5$ hours across varied drones, throttles, mic configurations, and environments, with SNRs ranging from $-57.2$ to $-2.5$ dB. The paper benchmarks noise suppression and audio classification methods (MVDR, spectral gating, MPSENet, and SSLAM), revealing fundamental limitations under extreme low-SNR conditions, especially for non-vocal human and ambient sounds, and provides concrete design guidelines for microphone placement, throttle strategies, and drone sizing. The dataset and findings offer a practical stepping stone for advancing drone audition technologies, enabling robust human-presence detection and informing hardware-software trade-offs for SAR operations.

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.

Figures (11)

• Figure 1: Figure (a) illustrates our experimental setup with the drone attached to a fixed aluminum frame, with two microphone arrays, $\mathbb{M}_{up}$ and $\mathbb{M}_{down}$, and a single microphone, $\mathbb{M}_{center}$. The source sounds (i.e., human vocal sounds, human presence sounds and ambient sounds) are transmitted through a speaker. (b) the actual setup -- the drone frame, microphone array, and the drones used.
• Figure 2: Figure (a) depicts the histogram of signal-to-noise ratios (SNRs) of all the data collected, and (b-g) depict the SNRs achieved across three different microphones and two drone throttle levels. Specifically, each box plot contrasts the SNR levels for different source loudness levels (60 dB and 90 dB), and microphone-drone distances (25 cm and 50 cm).
• Figure 3: Figure illustrates the current benchmarking pipeline, that takes as input the audio recordings from one of the microphones, $\mathbb{M}_{up}$, $\mathbb{M}_{center}$ or $\mathbb{M}_{down}$. Our pipeline consists of drone noise suppression (which we evaluate for three combinations of algorithms), followed by human-presence detection, that ultimately predicts audio file as corresponding to human sounds (vocal, non-vocal) or non-human sounds (ambient), or silence. Here $^*$ indicates the absence of beamforming for microphone, $\mathbb{M}_{center}$, as it consists of only a single microphone.
• Figure 4: SI-SDR scores for Hybrid noise suppression on HV sounds, for varied microphone types, drone-microphone distances, at (a) high, (b) low throttles. * indicate statistically significant difference$^{11}$ between pairs ($p<0.05$).
• Figure 5: SI-SDR scores for Hybrid noise suppression on HV sounds, for two drones, at two throttles, as well as drone-speaker distances. * indicate statistically significant difference$^{11}$ between pairs ($p<0.05$).