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

Authors: Maris Basha¹, Yuhang Wang¹, Xiaoran Chen², Longbiao Cheng¹, Luca Yapura¹, Anja T. Zai¹, Mathieu Salzmann², Richard Hahnloser¹
¹ Institute of Neuroinformatics, University of Zurich and ETH Zurich · ² Swiss Data Science Center and CVLab, EPFL
Paper: NeurIPS 2026 Evaluations and Datasets Track · Maintainer: Maris Basha (Institute of Neuroinformatics, UZH/ETH Zurich)
Provenance: recorded in the BirdPark aviary (Rüttimann et al., PeerJ 2025) at the Institute of Neuroinformatics; animal procedures approved by the Cantonal Veterinary Office of Zurich (license ZH148/2020).

Task. Given a detected zebra finch vocalization and the set of birds visible at that moment, determine which bird produced the call. Caller identity is verified physically from on-body accelerometer vibration; the accelerometer channel is withheld from benchmark models and appears only as a withheld on-body reference.

V33DA provides 33,625 verified vocalization events from 10 individually identified zebra finches across 3 experiments (2021–2023), each with synchronized 5-channel audio, multi-view video, calibrated 3D pose for every visible candidate, per-bird FM radio telemetry, and accelerometer-derived ground-truth labels.

	BP01	BP02	BP05	Total
Birds	4	3	3	10
Events	16,219	11,577	5,829	33,625
Recording days	10	3	3	16
Raw hours	~46h	~14h	~14h	~74h

Code: github.com/marisbasha/v33da  ·  Companion release: songbirdini/v33da_pp (V33DA++: overlap and longer-context files; not evaluated in the paper)

Quick start

pip install huggingface_hub
huggingface-cli download songbirdini/V33DA --repo-type dataset --local-dir data/v33da

Or in Python:

from huggingface_hub import snapshot_download
snapshot_download("songbirdini/V33DA", repo_type="dataset", local_dir="data/v33da")

Loading a sample

Each binary array column is serialized with numpy.save. To decode:

import io
import numpy as np
import pyarrow.parquet as pq

table = pq.read_table("data/v33da/v33da-00000.parquet")
row = table.to_pydict()

# 3D keypoints: (T_frames, N_birds, 5, 3) float32
kp3d = np.load(io.BytesIO(row["keypoints_3d"][0]))

# Audio path -> multichannel WAV
audio_path = row["audio_path"][0]  # e.g. "audio/juvExpBP01/2021-06-28/..."

The included explore.ipynb notebook shows how to inspect audio, video, pose, radio, and accelerometer data, and how to reproject 3D keypoints into 2D camera views using the shipped calibrations.

Modalities

Each parquet row is one vocalization event:

Column	Description
audio_path	5-channel cage-microphone WAV (24,414 Hz, float32); call onset to offset plus ±10 ms, 40–784 ms (mean 150 ms, median 125 ms)
accelerometer_path	Per-bird accelerometer WAV (verification signal; withheld from benchmark inputs)
video_path	Aligned composite MP4 clip (3 views, 47.68 fps)
keypoints_3d	Triangulated 3D pose, shape (T_frames, N_birds, 5, 3) float32: 5 keypoints per bird (beak, head, backpack, tail base, tail tip)
keypoints_2d_top	2D keypoints in top-view image space, shape (T_frames, N_birds, 5, 2) float32
keypoints_2d_back	2D keypoints in back-view image space, shape (T_frames, N_birds, 5, 2) float32
radio_*	21 per-bird BirdRadio signals per radio frame (see Radio telemetry)
vocalizer_idx	Ground-truth caller index
bird_color	Color identity of the vocalizer
experiment	Experiment name (juvExpBP01 / juvExpBP02 / juvExpBP05)
date	Recording date

Radio telemetry

Each bird's transmitter is received by four antennas (A–D); one radio frame spans 512 samples at 24,414 Hz, so radio frames coincide with video frames (47.68 Hz). Following BirdRadio's definitions (Rüttimann et al., PeerJ 2025):

Column	Quantity
radio_frq	PLL tracking frequency
radio_thetaS	master-vector phase angle $\theta_i$
radio_thetaA–radio_thetaD	per-antenna rotated-vector angles $\alpha_i^a$
radio_phiA–radio_phiD	per-antenna compensation phases $\phi_i^a$
radio_phiM	sum of the compensation phases
radio_powRA–radio_powRD, radio_powRM	baseband signal powers (dBFS)
radio_powNA–radio_powND, radio_powNM	baseband signal variances (dBFS)

BirdRadio does not compute an angle of arrival.

Ground-truth labels

Labels come from on-body accelerometer vibration, not from microphone-based localization. A WhisperSeg model proposes candidate windows from the demodulated signal; events are retained only when audible in microphones, showing characteristic on-body vibration, with no overlapping call from another bird. All candidates are manually reviewed.

Filtering

From 39,329 reviewed candidates:

• 1,768 rejected as invalid during manual review
• 3,784 removed because another bird vocalized during the same window (overlap filtering)
• 152 removed by 3D-to-2D reprojection error (> 40 px)

Yielding 33,625 released events.

Per-bird counts

Experiment	Bird	Events
BP01	blue	5,735
BP01	peach	2,277
BP01	red	3,963
BP01	white	4,244
BP02	blue	4,923
BP02	peach	4,589
BP02	red	2,065
BP05	brown	56
BP05	purple	2,140
BP05	yellow	3,633

Note on bird colors. Backpack color codes are experiment-local identifiers, not global bird IDs. A color that appears in two experiments (e.g. red in BP01 and BP02) refers to two different physical individuals; color-to-bird assignment is fixed within an experiment. All 10 released birds are distinct physical individuals; no bird participates in more than one experiment.

Release contents

Path	Description
v33da-*.parquet	23 pooled parquet shards
audio/	Multichannel microphone WAV files
accelerometer/	Multichannel accelerometer WAV files
clips/	Aligned composite MP4 clips
calibrations/	Per-experiment camera calibration files
explore.ipynb	Dataset exploration notebook
metadata.json	Release schema and filter metadata

Evaluation regimes

The benchmark code supports three evaluation regimes:

Regime	What transfers?	Use case
Session-disjoint	Same birds, different recording dates	Familiar-voice recognition
Held-out experiment	New cohort and recording period, same aviary	Callers never seen in training (cleanest unseen-caller test)
Leave-one-bird-out	One bird contributes no training calls (it may still appear as a silent candidate)	New caller under fixed sensing

Limitations

V33DA is collected in one aviary with one recording geometry across three experiments. Cross-experiment evaluation combines bird-identity shift, group-composition shift, and recording-date shift. The released benchmark excludes overlapping vocalizations and filters out samples with poor geometric consistency. Overlap-filtered calls and longer-context windows are released separately through the companion dataset songbirdini/v33da_pp, which is not evaluated in the paper. With ten birds and three experiments, held-out-experiment and leave-one-bird-out estimates are too noisy to rank the strongest methods against each other.

Citation

If you use V33DA, please cite:

@inproceedings{basha2026v33da,
  title     = {Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups},
  author    = {Basha, Maris and Wang, Yuhang and Chen, Xiaoran and Cheng, Longbiao and Yapura, Luca and Zai, Anja T. and Salzmann, Mathieu and Hahnloser, Richard},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
}