Two-stage acoustic detection and fine-grained classification of laying-hen vocalizations for heat-stress monitoring.

Heat stress is a critical environmental factor limiting the production and welfare of laying hens, yet continuous, non-contact monitoring at the flock level remains challenging. This study presents a heat stress assessment framework based on acoustic biomarkers, transitioning the monitoring emphasis from sporadic anomalous noises to detailed identification and intra-class analysis of typical vocalizations. Twenty healthy Jingfen 6 laying hens were randomly assigned to a control group (22 °C, 60% RH, n = 10) and a heat stress group (short-term heat exposure from 12:00 to 14:00 at 30-32 °C, 75-85% RH, THI ≥ 81, n = 10). The experiment was conducted over five consecutive days. Prior to the formal experiment, vocal activity rate (VAR) under normal temperature conditions was measured to characterize daily vocalization patterns, and a fixed analysis window (12:00 to 14:00) was subsequently selected. A two-stage automatic analysis pipeline was developed for continuous audio recordings. Vocal segments were detected by energy-based voice activity detection (VAD) and filtered by a binary classifier. The confirmed vocalizations were further classified into five categories by a dual-branch convolutional neural network (CNN) integrating basic acoustic features, with low-confidence outputs assigned to an "Others" category. The model achieved an accuracy of 97.58% (macro-F1 = 97.38%) for the binary classification task and 94.77% (macro-F1 = 94.54%) for the five-class task. Under heat stress, total vocal counts were higher during the first three days and declined during the final two days, while the control group remained stable. The proportion of vocalizations classified as "Others" was significantly higher in the heat stress group (P < 0.01), while the five predefined types showed no significant differences. Within-category acoustic analysis revealed shorter vocalization durations in the C1 and "Others" group (P < 0.05), and reduced pitch variability in C5 (F0 IQR, P < 0.05). These findings suggest that both vocal composition and intra-pattern acoustic features respond to thermal stress, providing preliminary evidence for the potential of structured acoustic monitoring as a non-contact approach for heat stress assessment in laying hens.