AI Paper Atlas

Research map / 3D vision / 3D generation

Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups

Maris Basha, Yuhang Wang, Xiaoran Chen, Longbiao Cheng, Luca N Yapura, Anja T Zai, Mathieu Salzmann, Richard Hahnloser

NeurIPS 2026 · 3D vision › 3D generation · approach: 2D-to-3D pretrained representations · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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 V33DA, a physically verified multimodal benchmark for vocal attribution (identifying which zebra finch produced a vocalization and where) in group settings.

Problem

Identifying which group member produced a vocalization and localizing it is central to social communication studies but is rarely evaluated with physically verified ground truth; existing resources either localize sounds without identifying the caller or recognize individuals from isolated vocalizat

Approach

A multimodal benchmark that jointly evaluates caller identification and spatial localization in zebra finch groups using physically verified ground truth and 3D candidate geometry.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use a slot-indexed identity output head (SELDnet++ slot)SELDnet++ · V33DA session-disjoint vs leave-one-bird-outtop-1 attribution accuracy (%): 86.5 -> 0.0Hurt
attribute through localization instead of directly with slot head (ablation)AudioANN · V33DA leave-one-bird-outtop-1 attribution accuracy (%): 0.0 -> 53.0Improved
remove microphones from the array (ablation)SpatialScorer · V33DA leave-one-bird-outtop-1 attribution accuracy (%): 51.2 -> 26.0Hurt
permute microphone channels (ablation)ACCDOA · V33DA session-disjointtop-1 attribution accuracy (%): 64.3 -> 38.8Hurt
zero audio at test time (ablation)ACCDOA and EINV2 · V33DA LOO / held-out BP01 / BP02accuracy change (points): -27.4 and -26.4 on LOOHurt
shuffle candidate geometry across events (ablation)ACCDOA and EINV2 · V33DA LOO / held-out BP01 / BP02accuracy change (points): -29.0 and -26.9 on LOOHurt
zero radio features at test time (ablation)ACCDOA and EINV2 · V33DA LOO / held-out BP01 / BP02accuracy change (points): +0.0No clear effect
zero radio features at test time (ablation)SpatialScorer · V33DA held-out BP05accuracy change (points): -14.8 on BP05; +6.3 on LOOMixed
add candidate conditioning without spatial reasoning (SELDnet++ open, PairRank, Whisper candidate, CST-former)candidate-scoring baselines · V33DA leave-one-bird-outtop-1 attribution accuracy (%): below 27.0 random referenceNo clear effect
smaller distance between vocalizer and nearest competing bird (below 10 cm vs above 50 cm)ACCDOA · V33DA LOOtop-1 attribution accuracy (%): 73.0 -> 46.6Hurt
physics-only SRP +-20 vs learned slot model under transferSRP +-20 · V33DA leave-one-bird-outtop-1 attribution accuracy (%): 0.0 -> 53.5Improved
per-experiment color gallery, outlier removal, occlusion filtering, 48x augmentation (ablation)K-NN backpack color classifier · 15 BirdPark experimentscolor accuracy (%): 88.9% -> 96.4%Improved

Experimental setup

Benchmarks: V33DA

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy