Research map / Video / Video-language understanding
Vocal Call Locator Benchmark (VCL) for localizing rodent vocalizations from multi-channel audio
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the VCL Benchmark, a large multi-channel audio and video dataset of 767,295 rodent vocalization events with 2D ground-truth source positions, and benchmarks a 1D-CNN localizer against the MUSE beamforming baseline.
Problem
Localizing animal-generated vocalizations in reverberant laboratory environments is poorly served by classical sound source localization methods, and no public datasets or benchmarks exist for rodent SSL.
Approach
A large-scale, ground-truth-labeled dataset of rodent sounds across three arenas (speaker, robot, earbud, solo and dyad recordings) plus two benchmark tasks (sound source localization with calibrated confidence sets, and vocalization attribution in dyads), with a raw-waveform 1D-CNN Gaussian-output baseline compared against MUSE.
Key results
- On Speaker-4M-E1 the DNN achieved a median error of 0.2 cm with 80.6% of predictions under 1 cm, versus 6.4 cm mean error for MUSE (5.9% under 1 cm).
- On Hexapod-8M-E2 the DNN mean error was 12.9 cm versus 18.1 cm for MUSE, showing DNNs outperform MUSE on all five single-source datasets but remain far from 1 cm resolution in harder settings.
- In dyadic attribution, 19-29% of calls were assigned to a single animal, while 65.2% (gerbil) and 71.7% (mouse) of calls had confidence sets containing both animals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DNN instead of MUSE | Sound source localization · Speaker-4M-E1 test set | mean error (cm): 6.4 -> 1.4 | Helped |
| Use DNN instead of MUSE | Sound source localization · Edison-4M-E1 test set | mean error (cm): 9.5 -> 1.4 | Helped |
| Use DNN instead of MUSE | Sound source localization · SoloGerbil-4M-E1 test set | mean error (cm): 14.3 -> 6.1 | Helped |
| Use DNN instead of MUSE | Sound source localization · Hexapod-8M-E2 test set | mean error (cm): 18.1 -> 12.9 | Helped |
| Use DNN instead of MUSE | Sound source localization · MouseEarbud-24M-E3 test set | mean error (cm): 11.3 -> 4.1 | Helped |
| Add temperature scaling (ablation) | DNN · Speaker-4M-E1 | ECE: 0.13 -> 0.05 | Helped |
| Add temperature scaling (ablation) | DNN · SoloGerbil-4M-E1 | ECE: 0.22 -> 0.05 | Helped |
| Use DNN instead of MUSE | Sound source localization · Speaker-4M-E1 | ECE: 0.17 -> 0.13 | Helped |
| Use DNN instead of MUSE | Sound source localization · Edison-4M-E1 | ECE: 0.12 -> 0.03 | Helped |
| Use DNN instead of MUSE | Sound source localization · SoloGerbil-4M-E1 | ECE: 0.24 -> 0.22 | Helped |
| Use DNN instead of MUSE | Sound source localization · Hexapod-8M-E2 | ECE: 0.22 -> 0.03 | Helped |
| Use DNN instead of MUSE | Sound source localization · MouseEarbud-24M-E3 | ECE: 0.25 -> 0.03 | Helped |
Surprising findings
- Deep networks trained on low-frequency gerbil calls failed to generalize to high-frequency sounds, and DNNs trained only on simulated room data barely exceeded chance on real data.
- MUSE baseline predictions were overconfident in reliability diagrams, while DNN confidence sets were also sometimes overconfident and required temperature scaling.
Limitations
- Only 2D localization from a single overhead camera; no 3D pose information.
- Covers only gerbils and mice in three arenas; other species such as marmosets, bats and birds are not included.
- Simulated data did not bridge the gap to real data, and dyadic attribution leaves most calls ambiguous.
Experimental setup
Datasets: Speaker-4M-E1 Edison-4M-E1 GerbilEarbud-4M-E1 SoloGerbil-4M-E1 DyadGerbil-4M-E1 Hexapod-8M-E2 MouseEarbud-24M-E3 SoloMouse-24M-E3 DyadMouse-24M-E3
Benchmarks: VCL Benchmark Task 1 (sound source localization) VCL Benchmark Task 2 (vocalization attribution)
Metrics: centimeter error (mean, median, % under 1 cm) expected calibration error maximal calibration error fraction of confidence sets containing 0, 1, or 2 animals
Baselines: MUSE
Builds on: MUSE image source method
Closest papers by meaning
- NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics (ICLR 2025 · similarity 0.88)
- OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations (NeurIPS 2026 · similarity 0.86)
- Spike-based Neuromorphic Model for Sound Source Localization (NeurIPS 2024 · similarity 0.85)
- BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics (ICLR 2025 · similarity 0.84)
- Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI models in Sound Localization (NeurIPS 2025 · similarity 0.84)
- SSLAM: Enhancing Self-Supervised Models with Audio Mixtures for Polyphonic Soundscapes (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.