WrenNet: Efficient On-Device Bird Audio NN

• WrenNet is an efficient neural network designed for real-time, on-device multi-species bird audio classification on severely resource-constrained microcontrollers.
• It employs a strictly causal, streaming architecture with a semi-learnable spectral feature extractor that blends logarithmic and linear frequency mappings to optimize avian signal processing.
• Demonstrating 70% test accuracy on a 70-species dataset and up to 90% on subsets, WrenNet efficiently meets strict energy and memory constraints for edge deployment.

Searching arXiv for the specified paper and closely related context. WrenNet is an efficient neural network for real-time, on-device multi-species bird audio classification on low-power microcontrollers, introduced for scalable biodiversity monitoring. It combines a streaming-compatible, strictly causal architecture with a semi-learnable spectral feature extractor that blends logarithmic and linear frequency mappings learned end-to-end to better match avian vocalizations. The system is designed for severely resource-constrained platforms, including devices with $\leq 1$ MB RAM and $\leq 100$ MHz processors, and is presented as the first practical framework for continuous, multi-species acoustic monitoring on low-power edge devices. On a 70-species dataset it attains 70.14\% test accuracy with a 136k-parameter model, and on an acoustically distinctive 8-species subset it reaches 90.76\% accuracy; on AudioMoth it consumes 0.077 J per 3 s inference (Ciapponi et al., 24 Sep 2025).

1. Problem setting and operational scope

WrenNet targets continuous, autonomous bird monitoring in remote habitats where power budget, memory budget, and deployment duration constrain model design more severely than in server or single-board-computer settings (Ciapponi et al., 24 Sep 2025). The motivating constraint is not merely low average power, but the requirement that inference run locally on microcontrollers with $\leq 1$ MB RAM and roughly 80 MHz-class processors, so that long retrieval cycles and cloud dependence can be avoided.

The system is framed around multi-species classification rather than the single-species or binary detection settings that have dominated prior embedded bioacoustic work. This distinction is operationally important: field campaigns benefit from local classification because it can support continuous monitoring and adaptive sensing without transmitting or storing all raw audio. At the same time, typical lab-grade models such as BirdNET are described as too heavy for microcontrollers, making architectural compression alone insufficient unless the entire inference pipeline is redesigned around causality, fixed-memory streaming, and integer deployment.

A common source of confusion is the phrase “multi-species.” In WrenNet, the task is multi-species classification at clip level, but the final student-training target per clip is single-label, consisting of one species or a no_bird class. The paper therefore addresses species inventory across clips rather than full polyphonic multi-label acoustic scene parsing. This distinction matters when interpreting results on overlapping vocalizations and acoustically dense field recordings.

2. Strictly causal architecture and streaming inference

The model architecture is explicitly streaming-compatible and strictly causal from frontend to classifier (Ciapponi et al., 24 Sep 2025). Input audio is processed as 3 s clips in 200 ms chunks at 32 kHz. The frontend computes an STFT with a 512-point FFT and a 320-sample hop, after which a 64-channel adaptive filterbank is applied. Temporal modeling is then handled without violating causality.

The convolutional encoder uses causal 1D convolutions with dilation and stride 1. Its blocks are PhiNet-inspired depthwise-separable convolution blocks, combining depthwise temporal filtering with pointwise channel mixing and squeeze-and-excitation channel attention. Hierarchical skip connections, described as MatchboxNet-like, retain both local and longer-range structure. The concrete configuration reported is a 3-layer MatchboxNet-style encoder with 32 base filters.

Temporal aggregation is performed by a unidirectional GRU with a 64-dimensional hidden state. Cross-frame state retention allows the GRU to preserve context across chunks, enabling continuous inference with fixed memory. This stateful design is central to the deployment strategy because it avoids the need to buffer large spectrogram tensors over long windows. A temporal attention module then selects discriminative frames before the final classifier outputs species probabilities; student training includes a no_bird class.

Two trained model scales are reported. The smaller full-dataset model has 57k parameters and the larger model has 136k parameters. MACs and FLOPs are not reported. The memory target is instead expressed through deployment feasibility under the AudioMoth constraint of $\leq 1$ MB RAM, achieved using CMSIS-NN deployment with external SRAM on an 8-bit bus. This suggests that intermediate activation management and state retention were treated as first-order design criteria rather than secondary implementation details.

3. Semi-learnable spectral feature extractor

A central contribution of WrenNet is a semi-learnable spectral feature extractor that replaces fixed mel scaling and also outperforms a fully learnable filterbank on the reported ablation (Ciapponi et al., 24 Sep 2025). The frontend begins with the standard STFT

$$X(k,n) = \sum_m x[m]\, w[m-nH]\, e^{-j 2\pi k m/N},$$

with spectrogram magnitude or power

$S(k,n) = |X(k,n)|^p.$

Filterbank energies are then computed as

$E_b(n) = \sum_k W_{k,b}\, |X(k,n)|^p,$

where the triangular weights $W_{k,b}$ are centered at warped frequencies.

The warping is defined over normalized frequency coordinates $x \in [0,1]$ using a convex combination of logarithmic and linear mappings:

$$f_{\log}(x) = f_{\min} \left(\frac{f_{\max}}{f_{\min}}\right)^x,$$

$\leq 100$0

A sigmoid gate interpolates between them,

$\leq 100$1

yielding

$\leq 100$2

The corresponding discrete FFT bin index is

$\leq 100$3

The learned parameters are the breakpoint $\leq 100$4, initialized at 8 kHz, and the transition width $\leq 100$5. The limiting behavior is explicit: $\leq 100$6 produces a hard switch at $\leq 100$7, whereas $\leq 100$8 yields a uniform blend of log and linear scaling. Because $\leq 100$9, $\leq 1$0, and $\leq 1$1 are monotonic, the convex combination preserves monotonic frequency order. Practical bounds are enforced through clipping in $\leq 1$2.

Training of the semi-learnable parameters uses enhanced learning rates—15$\leq 1$3 for the breakpoint and 5$\leq 1$4 for the transition width—together with gradient noise injection, oscillatory perturbations every 5 epochs, momentum resets, and validation-guided search within 10\% ranges. No explicit analytical $\leq 1$5 term is reported. Conventional log-compression, such as $\leq 1$6, is not explicitly described as used; the emphasis is on adaptive warping rather than post-filterbank compression.

4. Dataset construction, labeling, and optimization

The training corpus comprises 70 alpine bird species sourced from Xeno-Canto, with 150,645 downloaded files (Ciapponi et al., 24 Sep 2025). Recordings are resampled to 32 kHz mono and bandpass filtered from 150 Hz to 16 kHz. Segment extraction uses an amplitude envelope with a 50 ms sliding window and 10 ms hop, followed by prominence-based peak detection with a minimum inter-peak distance of 1 s. Top segments are selected via a 75th percentile threshold, and if no peak is found the first 3 s is used as a fallback. The resulting standardized 3 s clips total 150,557 segments.

The no_bird class is assembled from low-energy segments and selected ESC-50 ambient sounds, including rain, wind, fire, dog, cat, frog, footsteps, car horn, engine, and train, while excluding bird-related and urban classes for separation. The paper also describes the dataset as expert-informed and equipped with species-level difficulty annotations.

The supervision scheme combines curated hard labels with teacher soft targets from BirdNET-Analyzer using a confidence threshold of 0.05. The loss is an adaptive focal distillation objective,

$\leq 1$7

with $\leq 1$8, $\leq 1$9, and focal parameter $\leq 1$0, together with automatic class weighting for imbalance. For the single-label multi-class setting,

$\leq 1$1

and the distillation term is

$\leq 1$2

Optimization uses AdamW with learning rate $\leq 1$3, weight decay 0.01, cosine annealing, and up to 150 epochs. The schedule alternates cycles of joint training with network-only and filter-only epochs, concluding with joint refinement. Early stopping uses 35-epoch patience and batch size is 64. “Standard data augmentation” is reported, but specific augmentations are not enumerated.

5. Accuracy, subset behavior, and ablations

On the full 70-species task, the 57k-parameter model achieves 66.51\% test accuracy and 67.49\% F1, while the 136k-parameter model achieves 70.14\% test accuracy and 70.81\% F1 (Ciapponi et al., 24 Sep 2025). The corresponding learned spectral parameters shift from a breakpoint of 851 Hz with transition width 19 in the smaller model to a breakpoint of 1390 Hz with transition width 33.86 in the larger model. The reported capacity increase therefore improves both accuracy and F1 while also changing the learned spectral allocation.

Subset analyses expose where the architecture is strongest and where confusion persists. On an acoustically distinctive 8-species subset consisting of ravens, owls, herons, and woodpeckers, WrenNet reaches 90.76\% accuracy and 90.90\% F1; the reported Student Acc is 89.85\% and BirdNET Acc is 91.04\%. On a 13-species hard subset containing closely related thrushes, tits, and kinglets, accuracy is 77.47\% and F1 is 77.97\%, compared with Student Acc 73.90\% and BirdNET Acc 82.86\%. For the Regulus pair, the model reaches 85.64\% accuracy and 85.60\% F1, compared with Student Acc 83.60\% and BirdNET Acc 87.87\%.

Frequency-focused subsets show similar structure. High-frequency specialists reach 91.49\% accuracy and 91.55\% F1, with a learned breakpoint of 164 Hz and width 8.55; low-frequency species reach 91.63\% accuracy and 91.67\% F1, with breakpoint 237 Hz and width 9. In both cases the paper notes nearly linear spacing. This suggests that, for subsets dominated by narrower spectral regimes, the learned frontend can converge toward an approximately linear allocation of filter centers.

The frontend ablation on a 9-class subset of 8 birds plus no_bird is especially diagnostic. With 64 bins, fixed mel filtering yields 79.61\% test accuracy, fixed linear triangular filtering 81.45\%, the semi-learnable combined log-linear frontend 87.22\%, and a fully learnable filterbank 83.83\%. The direct comparison indicates that limited, structurally constrained adaptivity outperforms both fixed perceptual scaling and unrestricted filter learning in this setting. Most remaining errors occur among taxonomically and acoustically similar species, especially thrushes, tits, and kinglets.

6. Edge deployment, efficiency, limitations, and resources

Deployment is demonstrated on AudioMoth and Raspberry Pi 3B+, with the AudioMoth configuration described as a wildlife-monitoring microcontroller platform with an approximately 80 MHz processor, $\leq 1$4 MB RAM, and external SRAM on an 8-bit bus (Ciapponi et al., 24 Sep 2025). Audio is framed into 3 s clips and processed in 200 ms chunks; causal 1D CNNs and the unidirectional GRU retain cross-chunk hidden state, avoiding large spectrogram buffers and preserving a fixed memory footprint. On-device inference uses CMSIS-NN integer kernels on AudioMoth. The specific numeric precision is not explicitly reported, whereas BirdNET on Raspberry Pi 3B+ is 8-bit quantized and WrenNet on Raspberry Pi 3B+ uses TensorFlow Lite.

The reported runtime and energy values are summarized below. Energy is defined as

$\leq 1$5

over the inference window.

Platform / model	Time per 3 s input	Energy per inference
AudioMoth (WrenNet, CMSIS-NN)	1.69 s	0.077 J
RPi 3B+ (WrenNet, TFLite)	0.061 s	0.172 J
RPi 3B+ (BirdNET, 8-bit)	0.978 s	2.79 J

On Raspberry Pi 3B+, WrenNet is reported as approximately 16.2$\leq 1$6 less energy-intensive than BirdNET, based on $\leq 1$7, and approximately 16.0$\leq 1$8 faster, based on $\leq 1$9. On AudioMoth, average power during inference is 0.046 W, and the 1.69 s processing time per 3 s audio segment is described as compatible with continuous streaming because of strict causality and state retention.

Several limitations are explicit. Confusions among acoustically similar species persist; overlapping vocalizations and noisy environments remain challenging, especially with short 3 s clips and single-label targets; rare species with few samples and distribution shifts across regions or seasons are likely to affect generalization. Quantization ablations, compression nonlinearity comparisons, and robustness evaluations under noise, overlap, or domain shift are not reported beyond the subset analyses. Practical thresholding and post-processing specifics are also not detailed.

Future directions named in the paper include adaptive species-specific thresholds, on-device model updates, semi-supervised learning over unlabelled audio, larger curated datasets, and collaborative inference across distributed nodes. Reproducibility resources include a public repository with scripts for dataset creation, training, distillation, and model export at https://github.com/wren-framework/wrennet. The repository is described as providing training and export tooling, though explicit pre-trained weight files and licenses are not detailed.