A microphone on the airframe of a small multi-rotor UAV is dominated by rotor ego-noise, so spoken flight commands arrive at negative signal-to-noise ratio (SNR). We study small-footprint keyword spotting (KWS) for a ten-word command vocabulary under real ego-noise, training on one quadrotor and testing on another. Besides per-clip accuracy we measure the streaming false-alarm rate on 4.4 h of continuous rotor noise. We compare classical noise-robust front-ends (CMN, PCEN, spectral subtraction), test-time adaptation, and two front-ends that track the per-band ego-noise floor over the two seconds preceding the decision window and either subtract it (normalisation) or feed it to the network as a second input channel (conditioning). Per clip, all front-ends look alike: +2-3 points on average, up to +14 at -15 dB. On continuous rotor noise they differ sharply. Normalising front-ends fire about ten times more often than plain log-mel at the same threshold and end up below it at a budget of one false alarm per hour (-9 points at 0 dB). Conditioning keeps the baseline's false-alarm rate and turns its gain into detections (+8 points at -10 dB over three seeds); on top of PCEN it gives the best per-clip accuracy and false-alarm rate, and a level-anchored variant is also invariant to the microphone gain. Real drone+interferer recordings expose the remaining failure mode, environmental sounds and bystander speech, which training negatives halve. A single script reproduces all on-device numbers on the NVIDIA Jetson Orin NX flight computer, where the complete pipeline costs 3 ms per 100 ms hop on the CPU.
Figures & tables
Figure 1: On-board pipeline. Every 100 ms the last second of log-mel frames is combined with the noise-floor profile Nf tracked over the preceding two seconds (blue blocks are the proposed front-end) and classified; smoothed posteriors drive the flight controller over MAVLink.
Front-end
Train
−15
−10
−5
0
5
10
20
clean
mean
FAR
log-mel
–
15.8
23.3
38.9
61.2
73.2
80.0
85.6
98.0
54.0
100.0
log-mel
G
29.9
53.3
71.6
82.6
87.2
89.8
91.4
96.0
72.3
43.1
NFN
G
44.5
65.7
78.3
85.2
88.2
89.6
90.7
96.6
77.5
53.8
NPC
G
41.8
64.2
78.8
86.8
90.3
91.9
93.1
96.3
78.1
30.3
log-mel
E
42.8
69.9
83.5
89.6
92.8
94.6
95.7
96.7
81.2 ± 0.2
3.6 ± 2.1
CMN
E
55.3
73.2
84.0
89.6
92.6
94.2
95.3
97.0
83.5
7.3
Table 1: BC-ResNet-8 on the 12-class UAV command task, test words mixed with unseen-drone ego-noise: accuracy (%) per SNR, mean over the seven noisy SNRs, and false-alarm rate FAR (%) on noise-only input. Train noise: E = drone2 ego-noise, G = GSC noise, – = clean; ± = std over three seeds; bold = best in the E block.
Figure 2: Streaming operation (BC-ResNet-8, deployed decision logic): hit rate of commands embedded every 3 s in unseen-drone rotor noise vs. false alarms per hour on 4.4 h of speech-free rotor noise; one marker per threshold τ∈[0.5,0.95] (“0”: none in 4.4 h), curves for seed 0, error bars: std over three seeds at 1 FA/h. Open markers: normalising.
Figure 3: Stress tests (BC-ResNet-8 unless noted). (a) Accuracy at 0 dB SNR with the test input scaled by −20 to +20 dB. (b) Accuracy lost when the floor context comes from the other throttle level of the same drone (mean over both directions).
acc.
hit 0dB
FA/h @0.7
unk.
Front-end
mean
1 FA/h
rotor
non-sp.
speech
/100
log-mel
81.3
82.8
1.1
27
143
3.8
log-mel + neg.
81.6
88.7
0.0
15
66
6.2
NFN
83.0
75.4
9.8
40
124
5.2
NFN + neg.
82.8
60.7
12.5
24
93
6.2
NPC-a
84.0
86.1
0.9
42
132
4.5
Table 2: Beyond rotor noise (BC-ResNet-8 unless noted, unseen drone, τ=0.7 ): per-clip mean accuracy; streaming hit rate at 0 dB under 1 FA/h; false alarms per hour on rotor noise, on real drone+non-speech-interferer and drone+bystander-speech recordings; firings per 100 non-command words. “+ neg.”: ESC-50 events and overlapping non-command words as negatives.
Stage / model
Params
PC
Jetson
Jetson
CPU
CPU
TensorRT
Front-end (log-mel +Nf )
–
0.10
0.57
n/a
BC-ResNet-3 + NPC-a
57 k
1.07
2.64
1.32
BC-ResNet-8 + NPC-a
349 k
2.09
8.01
1.45
BC-ResNet-8 + PCEN+NPC-a
349 k
2.02
8.10
1.44
KWT-1 + NFN
558 k
1.56
6.61
0.75
Table 3: Per-hop (100 ms) processing time in ms, batch 1, median over 300–500 runs. PC: Intel i5-14600K, ONNX Runtime CPU, 2 threads. Jetson: Orin NX 16 GB, JetPack 6.2.1, MAXN; ONNX Runtime 1.24 CPU (2 threads) or TensorRT execution provider (FP16). The front-end always runs in NumPy on the CPU.
User-defined keyword spotting (UD-KWS) enables zero-shot wake-word detection from text, but existing systems learn speaker-invariant representations that cannot reject impostors uttering the correct keyword. We address this dual zero-shot setting -- unseen keywords and unseen speakers -- with ZP-KWS, a lightweight framework combining a phoneme-supervised audio encoder with a GE2E-pretrained compact speaker encoder (about 0.9M parameters). Multiplicative late fusion at inference grants each branch independent veto power, supporting modes from conventional detection to strict speaker-gated activation without retraining. On LibriPhrase, Google Speech Commands, and Qualcomm datasets, ZP-KWS reduces target-only FRR at 1% FAR by up to 60% relative to the strongest baseline while maintaining competitive keyword detection, all within a 1.55M parameter budget for edge deployment.
Ming-Hsiang Hu, Kuan-Tang Huang, Chien-Chun Wang +2
Dept. Computer Science and Information Engineering, National Taiwan Normal University, Taiwan · E.SUN Financial Holding Co., Ltd., Taiwan · United Link Co., Ltd., Taiwan
Passive acoustic sensing is an attractive modality for counter-unmanned aerial system (counter-UAS) defence: it is covert, low-cost, and effective against drones with small radar cross-sections or minimal radio emissions. We present EchoHawk, an open and fully reproducible reference pipeline that detects a drone from its rotor harmonics, estimates its blade-passing frequency, and localises it with a microphone array via classical wideband beamforming (delay-and-sum, MVDR, MUSIC) and time-delay processing (GCC-PHAT, SRP-PHAT), followed by temporal tracking. We evaluate the system on a physically transparent synthetic benchmark that pits drones against hard low-frequency harmonic confusers, such as ground vehicles, and on real recorded audio. Our central methodological contribution is a documented case of session-level data leakage in a widely used public dataset: because its recordings are pre-segmented into short clips, naive clip-level splits place adjacent slices of the same continuous recording in both training and test sets, inflating reported performance. Enforcing recording-session-grouped cross-validation reduces, for example, a random-forest baseline's detection probability at a 1% false-alarm rate from 0.796 to 0.745, yielding honest numbers. All code, figures, and a synthetic data generator are released so that every result runs without any download.
Microphones mounted on UAVs enable aerial acoustic scene analysis applications such as search-and-rescue, wildlife monitoring, and industrial inspection. However, drone rotor noise often dominates the mixture signal at SNRs well below -10 dB, making source recovery extremely challenging. Existing enhancement and source separation methods are typically designed for near-balanced mixtures and degrade substantially in drone audition settings. In this work, we propose DRONEAUDIONET, a drone noise suppression method that reframes a source separation model as a drone noise estimator. To better model drone-dominant mixtures, we introduce a learnable mask-scaling mechanism that allows mask magnitudes beyond unity, together with an additive residual correction term for improved drone estimation and source recovery. We train and evaluate our model on a publicly available drone audition dataset and test generalizability on an out-of-domain dataset with unseen drone hardware and flight modes. Results show that DRONEAUDIONET consistently improves downstream sound classification performance, with the largest gains observed for human vocal sounds. Our findings demonstrate the importance of drone-specific modeling for robust aerial acoustic perception and highlight the potential of source separation methods for real-world drone-assisted search-and-rescue.
Chitralekha Gupta, Soundarya Ramesh, Yifei Luo +1
School of Computing, National University of Singapore