Source-linked AI summary
Bad: Taming the Bioacoustic Data Deluge with a Bat Activity Detector
Stefano Ciapponi, Santiago Martinez Balvanera, Andrea Cesaretti, Elisabetta Farella, Kate E. Jones
TL;DR
Passive bat monitoring produces massive ultrasonic datasets, while legacy triggers struggle with acoustic confusers and deep models exceed microcontroller limits. BAD is a hardware-aware system-algorithm co-design that uses a lightweight INT8 CNN to gate ultrasonic streams on an edge microcontroller. In the prevalence regime, BAD rejects 99.41% of noise frames and delivers a > 33× precision gain over Goertzel filtering for long-term field deployment.
Problem
Passive bat monitoring produces massive ultrasonic datasets, while legacy triggers struggle with acoustic confusers and deep models exceed microcontroller limits.
Source: §1 · p. 1Method
BAD is a hardware-aware system-algorithm co-design that uses a lightweight INT8 CNN to gate ultrasonic streams on an edge microcontroller.
Source: §7 · p. 4Results
In the prevalence regime, BAD rejects 99.41% of noise frames and delivers a > 33× precision gain over Goertzel filtering for long-term field deployment.
Source: §6.2 · p. 4Takeaways & Limitations
BAD is reported to gate ultrasonic streams directly at the edge with all 14 layers offloaded to the EFM32PG26 MVP accelerator.
Source: §7 · p. 4In-situ field testing across natural ecosystems remains future work, pending availability of next-generation open-source hardware prototypes.
Source: §7 · p. 4
Abstract
from arXiv · showhide
Passive Acoustic Monitoring of bats generates massive ultrasonic datasets (>27 GB/night per node), straining edge storage and battery life. Legacy triggers fail against acoustic confusers, while deep models exceed microcontroller limits. We present a hardware-aware Bat Activity Detector (BAD) specifically designed to discriminate bat calls from hard biological and environmental confusers across variable sampling rates (192-384 kHz). Tailored for the Silicon Labs EFM32PG26 (MVP) in 8-bit integer precision, our model achieves 100 percent hardware offload across all 14 layers (17.2 KB Flash, 73.1 KB RAM). End-to-end preprocessing (74.00 ms for 76 frames) and inference (30.00 ms) of 100 ms clips at 192 kHz require 104.00 ms per clip. On spatially out-of-domain recordings under a realistic low-prevalence regime (r_pos = 0.05), BAD achieves an AUC-ROC of 0.9748 and suppresses 99.4% of non-target noise frames while retaining 65.3% of bat calls - delivering a >33x precision gain over classical Goertzel baselines.
1. INTRODUCTION
Ultrasonic bat monitoring produces large volumes of mostly non-target audio, while existing triggers struggle with outdoor confusers and the edge lacks a low-power detector for these signals. This work proposes a hardware-aware BAD co-designed with its deployment pipeline for constrained microcontrollers.
1. INTRODUCTION
- At 384 kHz and 16-bit resolution, one node generates over 27 GB per night, much of it without target activity or containing environmental noise.The passage also estimates over 8 TB of raw audio for a standard recorder deployed for a month.Source: §1 · p. 1
- Amplitude thresholds trigger on wind and rain, while single-bin Goertzel filters cannot distinguish frequency-modulated bat calls from ultrasonic confusers.The resulting noise recordings can drain battery and SD-card capacity, while soft calls may be missed.Source: §1 · p. 1
- The proposed low-power BAD uses hard negatives and can gate noise before SD-card writes or serve as a Stage-1 trigger for downstream classifiers.The passage names BatDetect2 as an example of a downstream classifier.Source: §1 · p. 1
- The implementation targets the Silicon Labs EFM32PG26 and uses DMA double-buffering and zero-copy threading for uninterrupted execution.The introduction positions the design for MCU architectures with dedicated matrix acceleration.Source: §1 · p. 1
- The work presents a system-algorithm co-design for hardware-accelerated microcontrollers and benchmarks window lengths and sampling rates against accuracy and RAM overhead.It also profiles inference latency and active power and energy consumption under continuous-deployment constraints.Source: §1 · pp. 1–2
2. BAT ECHOLOCATION CHARACTERISTICS
Bat calls are brief ultrasonic pulses with variable frequency structure and spacing. The paper uses variable-duration windows to capture calls and temporal context while limiting latency and memory demands.
2. BAT ECHOLOCATION CHARACTERISTICS
- Bat echolocation pulses typically last 1–20 ms, with search-phase inter-pulse intervals from 20 ms to over 100 ms.The passage cites for these ranges.Source: §2 · p. 2
- Call structures range from steep frequency-modulated sweeps to constant-frequency components across 10–120 kHz.Figure 1 is cited as an illustration of this range and variation.Source: §2 · p. 2
- Windows of 50, 100, and 200 ms are used to capture complete vocalization cycles and temporal context while keeping latency and memory double-buffering compact.The passage describes the windowing as accommodating transient FM chirps and sparse search-phase passes.Source: §2 · p. 2
3. DATASET CONSTRUCTION
The dataset pairs annotated bat vocalizations with background and hard-confuser negatives in balanced clips. Separate validation splits test performance on unseen audio from training sites and geographically distinct sites.
3. DATASET CONSTRUCTION
- Training uses uniform 50, 100, and 200 ms clips with a 1:1 positive-to-negative ratio and 10,000 clips per epoch.Positive clips are centered on BatDetect2 annotations with random temporal jitter of up to 50 ms.Source: §3 · p. 2
- Negative clips combine non-overlapping BatDetect2 background audio with hard negatives such as owl chicks, dormice, orthopterans, and friction from “Is That a Bat?”.The two negative sources contribute equal parts.Source: §3 · p. 2
- Validation uses an unseen-audio split from training sites and an out-of-domain split from geographically distinct sites with novel background noise profiles.Source: §3 · p. 2
4. PREPROCESSING & FEATURE EXTRACTION
The preprocessing adapts to multiple sampling rates and clip durations, then converts audio into cropped spectrogram features. Lightweight background tracking and spectral subtraction condition those features for inference.
4. PREPROCESSING & FEATURE EXTRACTION
- The feature pipeline adapts to sampling rates of 192, 256, or 384 kHz and clip durations of 50, 100, or 200 ms.Source: §4 · p. 2
- STFT uses a fixed 2 ms window and 50% overlap; the power spectrogram is cropped to 10–120 kHz and interpolated to 64 frequency bins.At 384 kHz, the FFT uses 768 samples; a nominal 200 ms input produces 134 time frames.Source: §4 · p. 2
- Figure 2 depicts preprocessed bat and non-bat 200 ms clips resampled at 256 kHz.Source: Figure 2 · p. 2
- An EMA background estimate with α = 0.4 is subtracted per spectral column to highlight transient pulses without PCEN’s compute and memory overhead.The EMA uses O(1) memory state overhead.Source: §4 · p. 2
5. MODEL ARCHITECTURE AND HW CO-DESIGN
BAD’s architecture is tailored to the EFM32PG26’s MVP accelerator, with an int8 network topology designed to keep all layers on hardware. Its convolutional backbone ends in an MVP-accelerated spatial reduction and a two-class classifier.
5. MODEL ARCHITECTURE AND HW CO-DESIGN
- The EFM32PG26 MVP natively accelerates int8 convolutions, depthwise convolutions, max pooling, spatial means, and dense layers.The supported operations shape BAD’s hardware-aware network design.Source: §5 · p. 2
- BAD’s topology enforces full layer offloading to the MVP rather than falling back to CPU software emulation.The described network uses a stem followed by depthwise-separable stages.Source: §5 · p. 2
- After progressively reducing spatial dimensions, an MVP-accelerated global mean reduction produces a 64-dimensional embedding for the two-class Bat vs. Non-Bat head.The final pointwise convolution expands the channel dimension to 64 before spatial reduction.Source: §5 · p. 3
- The model uses post-training int8 quantization and is trained for 20 epochs with Adam, cross-entropy, and label smoothing (ϵ = 0.1).Source: §5 · p. 3
6. EXPERIMENTAL EVALUATION & RESULTS
Across spatially out-of-domain recordings, BAD outperforms classical trigger baselines under low prevalence. Feature-window and sampling-rate choices reveal an accuracy–memory trade-off, while hardware profiling measures the costs of near-real-time operation.
6.1. Feature Parameter Exploration
- At 192 kHz, 50 ms clips peak at 0.7720 accuracy, 0.7311 F1, and 0.8076 AUC-ROC; extending to 100 ms raises AUC-ROC to 0.9599, with only marginal gains at 200 ms.The authors attribute the shorter-window deficit to insufficient temporal context; 100–200 ms windows capture full vocalization envelopes.Source: §6.1 · p. 3
- The 200 ms, 256 kHz configuration reaches AUC-ROC = 0.9710, while 100 ms at 192 kHz offers AUC-ROC = 0.9599 and AP = 0.9684 with half the frame memory footprint.Increasing the sample rate to 384 kHz does not monotonically improve accuracy.Source: §6.1 · p. 3
6.3. Hardware Profiling & Energy Evaluation
- A 200 ms window takes 200.63 ms end to end, while the proposed event-driven duty-cycled workflow remains planned for future field trials.The 200 ms configuration is reported with AUC-ROC = 0.9608.Source: §6.3 · p. 4
6.2. Trigger Comparison Benchmarks
- In the reported prevalence regime, BAD rejects 99.41% of noise frames and captures 65.30% of vocalizations, achieving 85.68% precision and an F0.5 of 80.65%, over 33× Goertzel’s precision.Table 2 compares performance on 100 ms clips under F0.5-optimized decision thresholds.Source: §6.2 · p. 4, Table 2 · p. 4
6.3. Hardware Profiling & Energy Evaluation
- For 100 ms clips at 192 kHz, 76-frame preprocessing takes 74.00 ms and inference 30.00 ms, for 104.00 ms end-to-end latency; the model uses 17.2 KB Flash and 73.1 KB RAM.Profiling found 100% hardware offload across 14 layers and 180.0 nWh (0.648 mJ) per inference.Source: Table 3 · p. 4, §6.3 · p. 4
7. CONCLUSION AND FUTURE WORKS
BAD is a hardware-accelerated edge detector that rejects non-target noise with high hardware offload and compact memory use. Its latency is near real time, while in-situ field testing remains future work.
7. CONCLUSION AND FUTURE WORKS
- BAD rejects 99.4% of non-target noise frames (AUC-ROC = 0.9748) and delivers a > 33× precision gain over Goertzel baselines on out-of-domain recordings at rpos = 0.05.The INT8 model offloads all 14 layers to the Silicon Labs EFM32PG26 MVP accelerator, using 17.2 KB Flash and 73.1 KB RAM.Source: §7 · p. 4
- 104.00 ms total latency for a 100 ms window (76 frames) combines 74.00 ms preprocessing and 30.00 ms inference.Execution uses zero-copy processing under DMA double-buffering.Source: §7 · p. 4
- In-situ field testing across natural ecosystems awaits next-generation open-source hardware prototypes.Planned trials will assess long-term battery performance, seasonal environmental robustness, and adaptive duty-cycling.Source: §7 · p. 4