Abstract
Monitoring biodiversity at scale is challenging. Detecting and identifying species in fine grained taxonomies requires highly accurate machine learning (ML) methods. Training such models requires large high quality data sets. And deploying these models to low power devices requires novel compression techniques and model architectures. While species classification methods have profited from novel data sets and advances in ML methods, in particular neural networks, deploying these state of the art models to low power devices remains difficult. Here we present a comprehensive empirical comparison of various tinyML neural network architectures and compression techniques for species classification. We focus on the example of bird song detection, more concretely a data set curated for studying the corn bunting bird species. The data set is released along with all code and experiments of this study. In our experiments we compare predictive performance, memory and time complexity of classical spectrogram based methods and recent approaches operating on raw audio signal. Our results indicate that individual bird species can be robustly detected with relatively simple architectures that can be readily deployed to low power devices.
Introduction
Traditional bioacoustic monitoring relies on battery-powered Autonomous Recording Units (ARUs) that continuously record audio to SD cards. This "dumb" recording strategy generates massive amounts of irrelevant data—silence, wind, or rain—causing storage to fill and batteries to drain within weeks. Consequently, researchers must frequently visit field sites to swap hardware, only to find that less than 10% of the data contains target species.
TinyChirp addresses this by bringing intelligent, on-device classification to low-power microcontrollers (MCUs). By identifying specific bird species in real-time, the system can selectively store only relevant audio, extending battery life from two weeks to nearly five months—enough to cover an entire breeding season on a single charge.
Methodology
The TinyChirp framework utilizes a two-stage binary classification pipeline. First, a lightweight Baseline Filter (a 9th-order Butterworth highpass filter) screens for signal power. If the signal exceeds a threshold, the system wakes a second-stage TinyML Inference model for rigorous classification.
The researchers evaluated both spectrogram-based (2D CNNs) and time-series-based (1D CNNs and Transformers) architectures. To overcome the memory constraints of MCUs (often <256KB RAM), the authors employed 8-bit Post-Training Quantization and a novel Partial Convolution technique.
Partial convolution allows the device to compute average pooling iteratively over a small sliding window. This strategy is crucial for fitting standard architectures like SqueezeNet into the limited RAM of edge devices without incurring Out-of-Memory (OOM) errors.
Experimental Results
The study utilized a curated dataset of corn bunting (Emberiza calandra) songs alongside environmental noise from Google AudioSet and non-target species from Xeno-Canto.
A key finding was the "Spectrogram Bottleneck": on microcontrollers, calculating Mel-spectrograms can take longer than the model inference itself (e.g., 2.4 seconds for a 3-second clip). Consequently, time-series models often provide better latency. Among the tested models, the Transformer-Time architecture emerged as a top performer, balancing high accuracy with efficient compute on the nRF52840 (ARM Cortex-M4) test board.
TinyChirp achieved a precision of over 0.98. By reducing SD card writes by 90%, the system extended the operational life of ARUs from 2 weeks to 18 weeks on a single charge, effectively enabling a "set it and forget it" approach for an entire season.
Conclusion and Future Work
TinyChirp demonstrates that robust bird species detection is possible on extremely constrained hardware. Future work will focus on scaling this binary classification to multi-species recognition and further optimizing architectures to avoid OOM limits on even smaller microcontrollers. This approach significantly lowers the logistical barriers to large-scale biodiversity monitoring.
Summary updated