Multichannel Alignment

Results

Task description

Task 1 challenges participants to synchronise pairs of temporally desynchronised multi-channel audio recordings. Given pairs of recordings from autonomous recording units (ARUs) or animal-borne bio-loggers and a small set of ground-truth synchronisation keypoints, participants design systems to estimate inter-channel time offsets. Submissions are evaluated by Mean Absolute Error (MAE) of the predicted keypoint timestamps, averaged across two data subsets: ARU recordings and zebra finch bio-logger recordings.

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 Guo, Xusen Yu, Chunrui Zhao, Shanzheng Guan and Gongping Huang
Wuhan University; Yichang Testing Technique R&D Institute
Confidence-routed hybrid GCC-PHAT and CV-Peak-BLSTM

Jury award

Ragib Amin Nihal, Benjamin Yen, Takeshi Ashizawa and Kazuhiro Nakadai
Institute of Science Tokyo
Clock-Drift-Constrained Structured Regression for Multi-Channel Alignment
Their end-to-end modelling approach is rigorous. The submitted models perform remarkably well on both data domains, showing that an unified approach can be used to solve both coarse and fine-grained alignment.

Submission ranking

Submissions are ranked on the average MAE (mean absolute error) across the two data subsets: ARU and zebra finch. Lower values indicate better performance.

Submission information Rank Test set
Corresponding Author System name Technical
report		 Official
rank		 MAE avg (ms) MSE avg MSE ARU MSE Zebra
finch		 MAE avg
(ms)		 MAE ARU
(ms)		 MAE Zebra
finch (ms)
Gongping Huang Confidence-routed hybrid GCC-PHAT and CV-Peak-BLSTM huangConfidenceRouted2026 1 309.68 0.3023 0.0001 0.6046 309.68 0.63 618.74
Gongping Huang Confidence-routed hybrid GCC-PHAT and CV-Peak-BLSTM huangConfidenceRouted2026 2 313.49 0.3055 0.0001 0.6110 313.49 0.63 626.36
Yuhuan You Cross-protocol audio-refined curve hybrid youCurveHybrid2026 3 330.37 0.2688 0.0217 0.5159 330.37 74.33 586.40
Yuhuan You Release-posgate audio-refined curve hybrid youCurveHybrid2026 4 334.10 0.2738 0.0217 0.5258 334.10 74.33 593.86
Yuhuan You Group-tail late-suffix curve hybrid prior youCurveHybrid2026 5 367.68 0.2740 0.0323 0.5156 367.68 150.94 584.42
Ragib Amin Nihal Clock-Drift-Constrained Structured Regression for Multi-Channel Alignment nihalStructuredRegression2026 6 368.14 0.2853 0.0224 0.5483 368.14 126.23 610.05
Ragib Amin Nihal Clock-Drift-Constrained Structured Regression for Multi-Channel Alignment nihalStructuredRegression2026 7 368.63 0.2827 0.0229 0.5426 368.63 133.25 604.01
Gongping Huang Confidence-routed hybrid GCC-PHAT and CV-Peak-BLSTM huangConfidenceRouted2026 8 369.82 0.3140 0.0171 0.6110 369.82 113.27 626.36
Aditya Bhattacharjee Deep Learning Baseline biodcase2026Task1 9 540.17 1.2374 0.0226 2.4522 540.17 116.71 963.64
Kouei Yamaoka AuxSync without SRO and selective refinement yamaokaAuxSync2026 10 865.59 3.2547 0.0000 6.5094 865.59 0.39 1730.79
Kouei Yamaoka AuxSync full yamaokaAuxSync2026 11 1016.23 3.6574 0.0002 7.3145 1016.23 0.93 2031.54
Kouei Yamaoka AuxSync full with outlier refinement yamaokaAuxSync2026 12 1078.28 3.9640 0.0000 7.9279 1078.28 0.39 2156.17
Aditya Bhattacharjee Nosync Baseline biodcase2026Task1 13 1171.94 2.8078 0.0161 5.5996 1171.94 110.12 2233.76
Aditya Bhattacharjee GCC-PHAT Baseline biodcase2026Task1 14 1242.40 3.8018 0.0180 7.5856 1242.40 120.42 2364.37
Colm Ennis Affine Sync ennisAffineSync2026 15 1458.88 4.7103 0.1077 9.3129 1458.88 275.43 2642.34

Technical reports

BioDCASE 2026 Task 1 Baselines

Bhattacharjee, Aditya
Queen Mary University of London

A Shared Recipe for Multichannel Bioacoustic Alignment

Ennis, Colm
Independent

Method summary
Features spectrogram
Embeddings BEATs embeddings
Training or augmentation [channel swap, sub-second keypoint sampling, endpoint trim]
Decision rule affine grid search

Confidence-routed hybrid GCC-PHAT and CV-Peak-BLSTM

Guo, Xin and Yu, Xusen and Zhao, Chunrui and Guan, Shanzheng and Huang, Gongping
Wuhan University; Yichang Testing Technique R&D Institute

Method summary
Features spectrogram
Decision rule confidence routing over GCC-PHAT and CV-Peak-BLSTM delay trajectories

Clock-Drift-Constrained Structured Regression for Multi-Channel Alignment

Nihal, Ragib Amin and Yen, Benjamin and Ashizawa, Takeshi and Nakadai, Kazuhiro
Institute of Science Tokyo

Method summary
Features learned embeddings
Embeddings BEATs
Training or augmentation synthetic-drift augmentation (zebra finch only; none on aru)
Decision rule structured affine-drift regression

AuxSync for Multi-Channel Audio Alignment

Yamaoka, Kouei
The University of Tokyo

Method summary
Features waveform cross-correlation
Embeddings n/a
Training or augmentation no
Decision rule global STO estimation by AuxTDE, global SRO estimation by DXCP/AuxTDE, and dataset-dependent local AuxTDE refinement

Curve Hybrid Priors for Multi-Channel Audio Alignment

You, Yuhuan
Peking University

Method summary
Features timestamp candidates and public annotation-derived offset statistics
Decision rule ARU: same-group previous-file median Channel-1 offset prior fitted from public ARU development annotations; zebra_finch: position-gated same-group blend prior: zebra-finch public-validation positions use a 75/25 tail9-public-curve and tail28-late-curve blend, while later suffix positions reduce to the tail28 affine-residual curve