Bioacoustics for Tiny Hardware

Results

Task description

Task 3 challenges participants to develop an automatic classifier of birdsong that complies with the resource constraints of low-cost, battery-powered autonomous recording units. Systems must classify vocalizations from 10 bird species on the ESP32-S3-Korvo-2 development board. Submissions are evaluated on classification performance, model size, inference time, and memory usage on a hidden test set.

Task description page

Awards

The official winning submission is determined by the official evaluation ranking. The Jury Award is selected by the task organisers based on the submitted systems and technical reports.

Official winning submission

Xin Hai, Menghang Yin, Xingyue Li, Jinzhe Wang, Gongping Huang, Yuzhu Wang
Wuhan University; Tampere University
WHU-IASP MobileNetV2-lite system 1

Jury award

Coming Soon!

Submission ranking

Submissions are ranked based on a composite score that balances classification accuracy on the hidden test set for both inference- and embedding model, and embedded performance metrics (model size, memory usage, and inference time) on the ESP32-S3 microcontroller. Higher rank score values indicate a better overall performance.

The rank score is computed as:

rank_score = (
    (accuracy_inference - 0.56) / (1 - 0.56) +
    (accuracy_tflite - 0.56) / (1 - 0.56) * 2 +
    (roc_auc_inference - 0.89) / (0.97 - 0.89) * 0.5 +
    (roc_auc_tflite - 0.89) / (0.97 - 0.89) * 0.5 +
    (4755032 - model_size_tflite_bytes) / (4755032 - 42232) * 0.5 +
    MAX((267313754 - macs_inference) / (267313754 - 23326923), 0) * 0.3 +
    MAX((567313754 - macs_tflite) / (567313754 - 3911888), 0) * 0.5 +
    MAX((477.46 - embedded_time_ms_preprocessing) / (477.46 - 241.12), 0) * 0.3 +
    MAX((2000 - embedded_time_ms_model) / (2000 - 287.53), 0) * 0.4 +
    MAX((2000 - embedded_time_ms_total) / (2000 - 287.53), 0) * 0.5 +
    MAX((1000000 - embedded_ram_usage_bytes) / (1000000 - 95900), 0) * 0.5
)

In case you have further questions regarding your submission or you want to discuss the challenge task with us, please feel free to use: BioDCASE-Tiny-2026 Discussion Thread.

Submission information Embedding Model (.tflite) Inference Model Embedded Performance on ESP32-S3
Rank Solution Label Abbreviation Solution Full Name Technical
Report 		Acc. ROC-AUC Model Size [Bytes] MACs Acc. ROC-AUC Model Size [Bytes] MACs RAM Usage [Bytes] Time Setup [ms] Time Preproc. [ms] Time Model [ms] Time Total [ms] Rank Score
1 Hai_task3_1 MNV2Lite-S1 WHU-IASP MobileNetV2-lite system 1 hai2026 0.7382 0.9578 78392 5382064 357292 N/A 95900 7.890 294.730 361.740 664.360 3.73
3 Hai_task3_2 MNV2Lite-HardFT WHU-IASP MobileNetV2-lite hard fine-tuned hai2026 0.7400 0.9503 78392 5382064 357292 N/A 95900 7.880 294.730 361.740 664.350 3.69
6 Hai_task3_3 DSCNN-s20k WHU-IASP DS-CNN conv-first s20k system hai2026 0.7218 0.9487 42232 3911888 213197 N/A 167948 5.060 308.560 287.530 601.150 3.59
4 Hai_task3_4 MNV2Lite-stemB WHU-IASP MobileNetV2-lite stemB hard fine-tuned hai2026 0.7345 0.9555 84320 5810224 379807 N/A 144044 7.670 300.310 437.290 745.270 3.63
8 guimaraes_task3_1 FreqTimeN FreqTimeNet (freq/time-separated CNN, PCEN + Perch distillation) guimaraes2026 0.7709 0.9559 156960 114917472 0.7782 0.9593 336408 114559499 N/A 3.38
11 Mannini_FBK_task3_1 MNNWAVE1 Waveform Frozen Frontend WrenNet mannini_fbk2026 0.6600 0.9254 4755032 324283506 0.6600 0.9254 346725 N/A 263528 21.530 477.460 3961.890 4460.880 1.75
5 You_PKU_task3_1 TVLow8423 TrainVal Mid-Feature Low-8423 Ensemble you_pku2026 0.7636 0.9655 894364 121567233 0.7636 0.9655 923225 119602017 2874184 6.360 248.380 28206.990 28461.730 3.61
2 You_PKU_task3_2 TV8414 TrainVal Mid-Feature Single BalancedDSCNN you_pku2026 0.7545 0.9581 293912 40522400 0.7545 0.9581 307962 39867339 2816872 2.900 241.120 9407.610 9651.630 3.70
7 You_PKU_task3_3 TVLow8424 TrainVal Mid-Feature Low-8424 Backup Ensemble you_pku2026 0.7636 0.9657 1192924 162089644 0.7636 0.9657 1232126 159469356 2885000 8.330 246.190 37603.960 37858.480 3.50
10 You_PKU_task3_4 CleanP14 Clean Validation Mid-Feature Pruned-14 Ensemble you_pku2026 0.7745 0.9677 4178756 567313754 0.7745 0.9677 4286089 558142746 N/A 2.49
9 Baseline Base Baseline Pytorch none 0.5628 0.8923 111408 23315232 0.5647 0.8921 393109 23326923 202556 4.440 413.630 330.360 743.990 2.61

Technical reports

BioDCASE Task 3: Efficient Bird Sound Recognition with Signal-processing Priors and Knowledge Distillation

Heitor R. Guimarães and Tiago H. Falk
INRS-EMT, Université du Québec

Hardware-aware MobileNetV2-lite with Teacher-pool Distillation for BioDCASE-tiny 2026 Task 3

Xin Hai and Menghang Yin and Xingyue Li and Jinzhe Wang and Gongping Huang and Yuzhu Wang
Wuhan University; Tampere University

WrenNet For BioDCASE-tiny 2026 Task 3

Leonardo Mannini and Stefano Ciapponi and Elisabetta Farella
Fondazione Bruno Kessler

BioDCASE 2026 Task 3 Technical Report: Mid-Frequency Log-Mel BalancedDSCNN Ensembles for Tiny Bird Audio Recognition

Yuhuan You
Peking University