Cattle / Mobile app, dataset and machine-learning methods review

CowRumIA: a cattle-audio app and a test-set caution

GitHub repository snapshot at pinned commit; Zenodo dataset v1.0.0 2026 · Note reviewed 2026-10-14

CowRumIA is a 2026 undergraduate project pairing a Flutter Android app with YAMNet embeddings and a dense classifier for four cattle vocalization categories. The Zenodo v1 record describes 624 labeled WAVs; the README reports 194 stressed beef, 39 calm beef, 350 stressed dairy and 41 calm dairy samples and advertises 85.6% accuracy on 125 test audios. In the pinned training script, a random stratified 80/20 utterance split creates X_test, then the same X_test and y_test are supplied to training as validation data while callbacks select epochs and adjust the learning rate; predictions are later evaluated on that same set. The reported score therefore is not an untouched final-test estimate. The size or direction of any resulting optimism cannot be determined from this review. This is a useful mobile-implementation comparator and a strong evaluation-design lesson, not evidence of animal communication or a validated welfare diagnostic.

What this does — and doesn’t — establish

The four output classes jointly encode cattle production type and stress/calm category, and the README class counts are imbalanced (350 in the largest class versus 39 in the smallest). The split code is utterance-level and does not group by animal, session or site; the reviewed materials do not establish whether repeated animals or recording conditions cross splits. Zenodo states expert validation but the reviewed record does not provide the annotation protocol, inter-rater agreement or independent operational welfare ground truth. The dataset contains mixed-source material, but upstream, file-level reuse terms are unresolved; Zenodo shows only a copyright notice. GitHub declares MIT for repository software, while the packaged TFLite assets and underlying YAMNet/model-weight provenance and terms were not separately verified. The README describes an intensity-weighted stress-level output and a confidence threshold, but the relationship to independently measured welfare severity is not established here. Its documented training command points to a file that returned 404 at the pinned commit. No effect size, clinical/welfare utility, device reliability, field performance or cross-site generalization was reproduced. These are limits of the reviewed evidence, not proof that the project authors have no additional records.

What we can build on

For mobile animal-audio tools, keep capture and review flows distinct from a diagnostic claim. Reserve validation data for model selection and protect a separate locked test set; group splits by the biological and recording units that could carry shared signal. Evaluate class-specific errors, calibration and simple context/background baselines, and define behavior labels with qualified domain reviewers. A phone interface and an accuracy number do not establish welfare value, message meaning or transfer to another species.

See the related tools and reading paths →

The next reading task

Keep the Zenodo audio and all model assets metadata-only. Before considering any dataset or model use, resolve dataset and upstream file-level rights, inspect the annotation protocol and available animal/session/site grouping fields, and obtain cattle-welfare and bioacoustics review. For a separately justified benchmark, define an independently observed target, compare majority/context/background/acoustic baselines, select group-held-out validation and a locked test, and report macro-F1, per-class recall, calibration and clustered uncertainty. Resolve the missing documented training path before claiming an exact reproduction. No Talk2Nature training or welfare-facing output is recommended from the materials reviewed.

Go to the sources

1. CowRumIA cattle-audio app and repository (pinned GitHub snapshot) ↗

   Pinned README and repository license describing a Flutter Android cattle-vocalization classifier, its reported results, class counts and listed model assets.

2. CowRumIA training split and evaluation script at pinned commit ↗

   Static review of the exact training file at the pinned CowRumIA GitHub commit, focusing on splitting, callbacks and final evaluation.

3. CowRumIA bovine vocalization dataset (Zenodo v1.0.0) ↗

   Official versioned Zenodo metadata for the CowRumIA cattle audio corpus, its described collection scope, file listing and rights statement.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.