acceptodds
Withdrawn from ICLR 2027

DCATVD: A Large-Scale Longitudinal Dataset for Analysis of Domestic Cat Vocal Development

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In longitudinal studies of animal vocalizations, machine learning models enable a deeper analysis and understanding of developmental patterns in animal communication and vocalizations. However, longitudinal studies remain constrained by the limited availability of high-quality, large-scale datasets. To address this gap in feline bioacoustics, this paper introduces the Domestic Cat Age Transition Vocalization Dataset (DCATVD), a large, robust dataset containing vocalizations from 135 individual domestic cats across 12 different breeds with precise metadata (exact birth-date, breed, age group, and individual cat ID). Our thorough analysis of this dataset reveals novel discoveries on how essential vocal variables develop across the lifespan of cats, such as shifts in dominant vocalization type from purr to meow and the increase of specific precise vocalization units (Elemental Cat Meow Units). This paper opens an avenue to advance bioacoustic representation learning through an accessible dataset and a reproducible pipeline.

Primary area:
datasets and benchmarks
Keywords:
animal language processing, bioacoustics, domestic cat vocalization dataset

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kPtkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

A dataset paper sources from youtube and instragram with life-long feline videos across multiple breeds built using a robust pipeline. There are some ethical concerns regarding the data licenses. Moreover, the evaluation has some problematic parts.

Critical strengths

This paper proposes a new dataset on cat vocalization sourced from different public videos. The main novelty of this dataset is the fact that it allows for longitudinal studies. The pipeline for the data processing looks robust with some of the steps verified by human annotators (so-called quality filters are in place).

Critical weaknesses

Classification pipeline: There are more recent animal sound encoders that may be used: perch2, avex family, birdnet, birdmae etc, although some are bird-dominated/focused they also work for non-birds as good encoders, particularly call types. AVES is already behind those other models in many benchmarks. The classification head is a GMM but linear probing, LSTM or even transformer heads may improve performance dramatically. Also it's unclear whether the last year is the best one for the task. I am also unsure about the scientific findings and the stats that were used to validate them (the entire clustering part).

Questions for authors

Has this been validated? "We eliminated any video with a confidence score of under 0.50 using the prompt “image of a cat.”" How do you account for multiple cats in the same video? How do you account for cats that don't get to live in the entire expected life-span? What is their percentage if so? Not all of these classes are vocalizations because some of the sounds are not produced by the vocal tract: "The model classifies cat vocalizations into eight broad types: chatter, growl, hiss/spit, meow, purr,pain cry, trill, and yowl/howl" ie hiss, purr It is unclear whether there is a manual test set for call types. I think it should be and not semi-automatic. I am unsure about the clustering evaluation. I think the repertoire of a cat sits on a continuum and the clusters and number of clusters also change with age and are quite subjective, maybe report clustering metrics (nmi, ari) against human annotated classes. I am unsure whether the authors use the correct statistical tests. These tests come with some assumption on data and feature correlation and they should not be used blindly (explain why you use a test parametric vs non-parametric)

Reviewer_6hFAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors present a cat vocalization dataset consisting of 9k vocalizations from 135 unique cats across 12 breeds (~4.4 hours of vocalization). Their dataset creation pipeline consists of identifying cat channels on YouTube and Instagram, pulling cat metadata including age and breed information, pulling audio and extracting cat vocalizations, classifying them into types, and extracting information from 25ms contiguous chunks of vocalization. Many quality control checks are performed along the way and code is provided. Lastly, the authors present several analyses of the collected data including separating call frequency by life stage both across and within breed, clustering the 25ms chunks and also plotting these by life stage.

Critical strengths

A large amount of work has gone into making this dataset. In particular, the data quality checks seem to be rigorous and well-executed.

Critical weaknesses

I believe its utility of this dataset to animal vocalization researchers to be fairly limited. While the total duration of vocalizations is large, the number of cats within each breed or life stage is limited, which I suspect would make many fine-grained analyses underpowered given the data. Because of this, I expect it will have limited impact.

Another substantial issue is the low level of novelty given the largely overlapping methodology of Lekhak et al. (2025), a dog dataset method.

I think the clustering analysis of the ECMUs (25ms chunks) is a reasonable quality assurance step to identify and remove noise. It could also be useful as a preprocessing step to tame the complexity of the audio data, but I would not consider it proof that the data is in fact clustered, and not distributed along a continuum.

Questions for authors
  1. The dataset is described as longitudinal. Are the same cats recorded over two or three age categories? What are the age ranges from first to last video?
Reviewer_UbmHOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces the Domestic Cat Age Transition Vocalization Dataset (DCATVD) to address the lack of high-quality and large-scale datasets for bioacoustic research on domestic cats. The dataset is constructed by collecting cat videos from YouTube and Instagram and processing them through four stages: Channel Discovery and Selection, Video Filtration and Age Determination, Audio Preprocessing, and Meow Unit Segmentation. Furthermore, this work carries out age-related acoustic analyses using the proposed dataset.

Critical strengths

The dataset construction has certain value: the meows extracted in DCATVD are associated with metadata on birth date, breed, and age group, providing a unified data foundation for studying age-related acoustic changes in cats.

Critical weaknesses
  1. Dataset quality issues: 1) The data are primarily sourced from social media videos, and this collection method does not guarantee that the data adhere to the longitudinal concept emphasized in the paper. 2) The 9,135 MUs are derived from 135 cats, and the number of vocalization segments should not be equated with the number of independent biological samples, particularly when a small number of cats contribute a large number of segments. 3) The age distribution is highly imbalanced.
  2. Insufficient verification: The main contribution of this paper is the dataset, yet it lacks sufficient experiments on downstream tasks to evaluate the practical utility of the proposed dataset. This falls short of the acceptance criteria for the ICLR main conference.
Questions for authors
  1. What metrics or analyses support the longitudinal concept emphasized in the paper?
  2. Are the training and test sets split at the individual-cat level, or can the same cat appear in both sets?
  3. Do ECMUs exhibit cross-individual stability, and how many distinct cats are represented in each ECMU category?

Comments on the paper

1 comment

We thank the reviewers for their time. We are withdrawing the paper to pursue a venue better suited to the work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.