Multidomain feature fusion for recognizing chicken vocalizations under heat stress

Abstract

Heat stress is a key factor influencing the production performance and welfare of chickens. Chicken vocalizations reflect their physiological and health status. Whether chickens are experiencing heat stress can be assessed by analyzing their vocalizations. In this study, a lightweight chicken vocalization classification-fusion (CVC-Fusion) model based on multidomain feature fusion was proposed to recognize chicken vocalizations under heat stress and thermoneutral conditions. CVC-Fusion uses two parallel feature extraction branches to model one-dimensional frequency-domain features and two-dimensional spectrogram features of chicken vocalizations. Time- and frequency-domain features were calculated from laboratory recordings collected at three rearing densities and five age groups to characterize differences in acoustic features under heat stress and thermoneutral conditions. An augmented laboratory chicken vocalization dataset was constructed to train and evaluate CVC-Fusion. Additionally, a real-world chicken vocalization dataset collected from a semiopen cage-free chicken house was used for transfer learning and performance evaluation. The results revealed that heat-stress vocalizations generally exhibited greater energy and more concentrated spectral distributions than thermoneutral vocalizations did. In the laboratory environment, CVC-Fusion achieved $accuracy$, $precision$, $recall$, and $F_1$ values of 96.86 %, 97.03 %, 96.71 %, and 96.87 %, respectively, with a weighted Cohen’s kappa of 0.937 and an area under the receiver operating characteristic curve (ROC-AUC) of 0.986, using only 0.88 million parameters. Grad-CAM visualization verified the effectiveness of the dual-encoder architecture of CVC-Fusion in capturing key features and achieving information complementarity. After transfer learning, CVC-Fusion achieved $accuracy$, $precision$, $recall$, and $F_1$ values of 80.69 %, 80.70 %, 80.69 %, and 80.62 %, respectively, with a weighted Cohen’s kappa of 0.607 and an ROC-AUC of 0.794 in a real-world environment. The results demonstrate the effectiveness of CVC-Fusion for recognizing chicken vocalizations under controlled laboratory conditions while highlighting that accurate heat-stress recognition using a single audio modality remains challenging in semiopen cage-free chicken houses.

Introduction

The broiler chicken industry, as a vital component of the poultry sector, plays a crucial role in ensuring the supply of poultry meat and promoting agricultural economic development (China Animal Agriculture Association, 2022). With global average temperatures increasing annually amid the intensifying effects of global warming, extreme heat events have become increasingly frequent (IPCC, 2023). Consequently, heat stress in broiler chickens poses an increasingly severe threat to the farming industry. In China, broiler chickens are primarily raised in semiopen cage-free houses. While this farming model offers advantages such as a high rearing density and low operational costs (Fan et al., 2022), effective environmental controls are lacking. During summer, the temperature distribution within these houses becomes uneven, with the temperature in certain areas exceeding the optimal growth value for broiler chickens. Moreover, chickens have a high metabolism and generate significant amounts of heat, which, coupled with these other factors, contributes to heat stress (Slawinska et al., 2020), negatively affecting production performance and welfare levels (Shakeri et al., 2019). Currently, the assessment of heat stress in chickens relies primarily on inspections by farm workers. However, this method is subjective, labor intensive, and inefficient (He et al., 2022). With the aging population and rising labor costs, an automated and intelligent method is needed to replace manual inspections and support the advancement of modern poultry farming (Zhao, 2019).

Chicken vocalizations convey various information that reflects their health, emotions, and behaviors (Briefer and Comber, 2012; Lv et al., 2023). Acoustic analysis technology, which is efficient, noninvasive, and cost effective, offers significant potential for identifying vocalizations associated with heat stress. Through acoustic analysis, the underlying causes of specific vocalizations, such as the repeated, high-energy, and relatively loud calls made by chickens when they are experiencing distress, can be identified (Herborn et al., 2020).

In recent years, livestock and poultry vocalization analysis has been extensively investigated, primarily focusing on stress recognition and respiratory disease detection. With respect to stress recognition, certain vocalizations, such as alarm calls, squawks, and gakel sounds, from hens were significantly correlated with environmental temperature and humidity indices (Du et al., 2020). Additionally, turkeys under heat stress conditions produced 43 % more vocalizations than those under thermoneutral conditions did (Liu et al., 2018). Moreover, stress vocalizations in piglets, such as those induced by hunger, cold, or pain, were identified using a decision tree algorithm (Cordeiro et al., 2013). With respect to respiratory disease detection, data mining methods have been employed to extract effective time-domain and frequency-domain features from livestock and poultry sounds, enabling the identification of calf coughs or sneezing sounds associated with diseases such as Newcastle disease, bronchitis, and avian influenza (Banakar et al., 2016; Carpentier et al., 2019; Sun et al., 2024).

However, the machine learning methods employed in the above studies generally relied on manually extracted features, which are limited by the subjectivity and expertise of researchers (Du et al., 2021). In contrast, deep learning methods, with their advantage of end-to-end modeling, offer higher accuracy and reliability in bioacoustic detection scenarios (Stowell et al., 2019; Xie et al., 2024). The input to deep learning models typically consists of 1D frequency-domain data or 2D spectrograms of livestock and poultry sounds. High-dimensional representations of sound signals are obtained through convolutional operations or multihead self-attention mechanisms and are subsequently processed by classification heads to recognize specific vocalizations (Lev-ron et al., 2024; Mao et al., 2022; Cuan et al., 2020). For example, a convolutional neural network (CNN) was trained on log-mel spectrograms of chicken vocalizations to classify laying hen calls, rooster crowing, and background noise (Lin et al., 2025). Additionally, a CNN was employed to automatically recognize distress calls, short peeps, warbles, pleasure notes, and other vocalizations, revealing that chicken vocalizations are strongly influenced by circadian rhythms and bird age (Carvalho et al., 2025). These methods have demonstrated competitive recognition performance in specific tasks. However, the potential complementary relationships between multidomain features remain insufficiently explored, which limits the ability of existing models to enhance feature representations and effectively distinguish between different sound categories. Furthermore, while previous research has addressed various types of stress in livestock and poultry, studies focusing specifically on vocalizations under heat stress conditions remain scarce. In particular, studies on the differences between vocalizations under heat stress and thermoneutral conditions are lacking.

The general objective of this study was to develop a multidomain feature fusion model for recognizing chicken vocalizations under heat stress conditions. The primary contributions of this study are summarized as follows:

• (1)An original laboratory chicken vocalization dataset comprising three rearing densities and five age groups was established to characterize differences in time- and frequency-domain features between vocalizations recorded under heat stress and thermoneutral conditions.
• (2)A lightweight chicken vocalization classification-fusion (CVC-Fusion) model based on multidomain feature fusion was developed to recognize chicken vocalizations under heat stress conditions.
• (3)The contributions of these multidomain features to the model’s decision-making process were investigated through ablation experiments and feature visualization.
• (4)A real-world chicken vocalization dataset collected in a semiopen cage-free chicken house was constructed to evaluate the performance of the CVC-Fusion model after transfer learning.

Materials and methods

Experimental setup and data acquisition

The study involved 320 male chickens of the Youhuang 5B breed. All the chickens were purchased at 20 days of age from Jiangsu Lihua Co., Ltd., in 2024 and subsequently housed in a room (Room 0) equipped with a comprehensive environmental control, feeding, and lighting system. They were raised at a density of 8 birds/m². All experimental procedures were conducted following the Care and Use of Animals Guidelines of Zhejiang University (Hangzhou, China), and the protocol was approved by the Committee on the Ethics of Animal Experiments of Zhejiang University. As shown in Fig. 1(a), the experiments were conducted in six identical rooms (Rooms 1–6), each measuring 3.2 m in length, 3.0 m in width, and 2.6 m in height. Each room was equipped with heating panels, air conditioning, and temperature and humidity sensors, with a 5 cm layer of bedding on the floor. The light was on from 6:00 to 22:00 and controlled by a timer, with an intensity of 25 lux. A microphone (Earthworks M50, Earthworks Inc., USA) was positioned 1.8 m above the floor to capture indoor sound signals at a sampling rate of 44.1 kHz; the microphone was connected to an audio interface (Zen GO Synergy Core, Antelope Audio, Bulgaria). A camera (DS-2TD2636B-10/P (B), Hikvision, Hangzhou, China) was mounted 50 cm above the floor on one side of the room to capture video data of the chickens, with a resolution of 2688 × 1520 pixels and a frame rate of 25 fps. Both the camera and the audio interface were connected to a computer for data acquisition and storage. The chickens were acclimated in a central area of 1.0 m by 1.0 m within each room, where feeders and waterers were provided. To collect vocalization data from chickens of different ages, 4 chickens from Room 0, with ages 21, 28, 35, 42, and 49 days, were transferred to Rooms 1 and 2 for three days of rearing at a density of 4 birds/m². Eight chickens were moved to Rooms 3 and 4 for three days at a density of 8 birds/m², and 12 chickens were moved to Rooms 5 and 6 for three days at a density of 12 birds/m². This arrangement ensured that the chickens in each room adapted to the new environment.

Fig. 1.

Schematic diagram of the experimental setup in the laboratory environment for collecting sound and video data, along with example of the collected RGB image. (a) Schematic diagram of the experimental setup; (b) Example of the collected RGB image.

The experiments were conducted when the chickens in each room were 24, 31, 38, 45, and 52 days old. Between 12:00 and 16:30, Rooms 1, 3, and 5 were heated to 35 ± 0.5°C, whereas Rooms 2, 4, and 6 were maintained at a constant temperature of 23 ± 0.5°C. All other conditions in the rooms were kept the same, as shown in Table 1. During the experiment, continuous sound and video data were collected using microphones and cameras; thus, vocalization data were collected under heat stress and thermoneutral conditions at three densities (4, 8, and 12 birds/m²) for chickens at 4, 5, 6, 7, and 8 weeks of age. An example of the collected RGB images of the chickens is shown in Fig. 1(b).

Table 1.

Temperature settings under three rearing densities and five age groups in the experiment.

Groups	Room	Temperature (°C)	Rearing density (birds/m²)	Weeks of age (w)
Experimental Group	Room 1	35 ± 0.5	4	4, 5, 6, 7, 8
	Room 3		8
	Room 5		12
Control Group	Room 2	23 ± 0.5	4
	Room 4		8
	Room 6		12

To collect vocalization data from chickens in a real-world farming environment, an additional experiment was conducted in the summer of 2025 at a semiopen cage-free chicken house in Hefei city, Anhui Province, China, as shown in Fig. 2. The experimental subjects were commercial yellow-feathered broiler chickens, which were 49 days of age at the start of the experiment. The experiment was performed over two weeks. The camera used in the aforementioned experiment and a microphone (DS-2FP4021-B; Hikvision, China) were connected via an audio interface to continuously record video and audio data inside the chicken house. The video sampling rate and frame rate remained unchanged, while the audio sampling rate was set to 32 kHz. This approach enabled the acquisition of RGB video data of the chickens and acoustic data from the real-world chicken house.

Fig. 2.

Example of an RGB image of chickens in the real-world farming environment.

Overall methodological framework for chicken vocalization recognition

The complete methodological workflow is summarized in Fig. 3. Audio and video data were first collected in controlled laboratory rooms and the semiopen cage-free chicken house. Discrete and complete chicken vocalization events were segmented from the continuous recordings, cross-checked against the synchronized video data, and assigned group-level heat-stress or thermoneutral labels. The segmented vocalization samples were framed, windowed, and transformed using the short-time Fourier transform (STFT) to generate paired frequency-domain signals and spectrograms (Portnoff, 1981). For the laboratory data, augmentation strategies were applied to construct the augmented laboratory chicken vocalization dataset, whereas the processed recordings from the semiopen chicken house constituted the real-world chicken vocalization dataset. The augmented laboratory chicken vocalization dataset was used to train the chicken vocalization recognition model and evaluate its performance under laboratory conditions. The real-world chicken vocalization dataset was then used to fine-tune the laboratory-trained chicken vocalization recognition model through transfer learning, and its performance under real-world farming conditions was evaluated. The detailed procedures for each stage are described in the following sections.

Fig. 3.

Methodological workflow for chicken vocalization recognition in laboratory and real-world environments.

Audio preprocessing methods

With respect to the continuous audio files obtained in this experiment, the increase in the indoor temperature in the experimental group to a stable value of 35 ± 0.5°C took approximately 30 min. Therefore, only the audio from 12:30 to 16:30 was retained for subsequent processing. The audio contained numerous segments of silence and noise. In this study, Audacity 3.7.0 software (https://www.audacityteam.org/) was used to manually segment discrete and complete chicken vocalization events from the continuous audio signals. Additionally, the timestamp was cross-referenced with the video data to ensure that the vocalizations were not caused by chicken fights or sudden scares, which would result in “screaming” sounds. This procedure yielded two categories of chicken vocalization samples: (1) vocalizations recorded when most of the chickens exhibited panting behavior under heat stress and (2) vocalizations recorded when most of the chickens displayed normal behavior without typical stress responses. With respect to the laboratory audio data, 874 samples were obtained when most of the chickens exhibited heat-stress panting behavior, and 1,097 samples were obtained when most chickens exhibited normal behavior, with audio durations ranging from 0.30 to 1.32 s. These samples constituted the original laboratory chicken vocalization dataset. With respect to the real-world audio data collected in the farming environment, 1,323 samples were obtained when most of the chickens exhibited heat-stress panting behavior, and 1,004 samples were obtained when most chickens exhibited normal behavior, with audio durations ranging from 0.46 to 1.66 s. These samples constituted the real-world chicken vocalization dataset.

As chicken vocalizations are nonstationary signals, the signals were first framed into multiple short-time segments; each segment was considered a stationary signal, thus allowing for STFT analysis. The frame length was set to 2048 samples, with a frame shift of 512 samples. The signals were then windowed to reduce the spectral leakage caused by signal truncation, improving the spectral smoothness. A Hanning window was used in this study, which was calculated as shown in Eqn. (1).

$$w\left(n\right)=0.5\left(1-\text{cos}\left(\frac{2\pi n}{N-1}\right)\right)$$ (1)

where N is the window length and n is the current sample index, with n satisfying $0\le n\le N-1$.

In the conversion of the chicken vocalization audio data into corresponding frequency-domain features and spectrograms, only the data in the 1000–5500 Hz range were retained, as the effective chicken vocalization information was within this frequency range (Section 3.1). This approach also helped prevent interference from noise outside this range, particularly low-frequency noise, during model training.

Data augmentation methods

In this study, various data augmentation methods, including time scaling, time shifting, Gaussian noise addition, and random cropping and padding, were employed to improve the generalizability of the chicken vocalization recognition model (Mao et al., 2022). These augmentation methods were applied to the raw audio data. Specifically, time scaling was performed to simulate changes in vocal frequency by speeding up or slowing down the audio playback, whereas time shifting was performed to simulate real-world recording device delays and audio alignment errors by altering the starting point of the audio. Mild random Gaussian noise was added to the original audio to increase the robustness of the model to background noise interference, and random cropping and padding were applied to the chicken vocalization audio to improve the adaptability of the model to input data with varying sequence lengths.

Methods for calculating audio time-domain and frequency-domain features

To analyze the differences in chicken vocalizations under heat stress and thermoneutral conditions, both time-domain features, including the short-time energy (STE) and zero-crossing rate (ZCR) (Kedem, 1986; Xie and Zhu, 2019), and frequency-domain features, including the spectral centroid, spectral flatness, spectral contrast, and spectral flux (McKinney et al., 2003), were computed. The frequency-domain features were calculated after the original signal was framed, windowed, and transformed into the frequency domain using the STFT. The feature calculations included the following: 1) the mean time-domain and frequency-domain features for all the chicken vocalization samples collected for each age group across the three densities, which were used to analyze the overall time-domain and frequency-domain features of the chicken vocalizations at different ages, and 2) the mean time-domain and frequency-domain features for all the chicken vocalization samples collected at each density across the five age groups, which were used to analyze the overall time-domain and frequency-domain features of the chicken vocalizations at different densities.

The STE is used to measure the variation in the power intensity of an audio signal within a short time frame and is calculated as shown in Eqn. (2).

$$E\left(n\right)=\sum _{m=-\infty }^{\infty }{\left[x\left(m\right)w\left(n-m\right)\right]}^2$$ (2)

where $E\left(n\right)$ is the short-time energy of the n-th frame, $x\left(m\right)$ is the amplitude of the audio signal at time point m, and $w\left(n-m\right)$ is the window function centered at time point n.

The ZCR represents the number of times that the signal amplitude crosses zero within a given time interval and is calculated as shown in Eqns. (3) and (4).

$$Z_n=\frac{1}{2N}\sum _{m=0}^{N-1}\left|sgn\left(x\left(m\right)\right)-sgn\left(x\left(m-1\right)\right)\right|$$ (3)

$$sgn\left(x\left(m\right)\right)=\{\begin{array}{c}1,x\left(m\right)\ge 0\\ 0,x\left(m\right)<0\end{array}$$ (4)

where $x\left(m\right)$ represents the amplitude of the audio signal at time point m.

The spectral centroid represents the center of mass of the frequency components of the signal and is calculated as shown in Eqn. (5).

$$Centroid=\frac{{\sum }_{k=0}^{N-1}{f}_k\left|X\left(k\right)\right|}{{\sum }_{k=0}^{N-1}\left|X\left(k\right)\right|}$$ (5)

where $f_k$ is the actual frequency corresponding to the k-th frequency component and $\left|X\left(k\right)\right|$ is the amplitude of the k-th spectral component.

The spectral flatness is the ratio of the geometric mean to the arithmetic mean of the power spectrum and is calculated as shown in Eqn. (6).

$$Flatness=\frac{{\left[{\prod }_{k=0}^{N-1}X\left(k\right)\right]}^{1/N}}{\frac{1}{N}{\sum }_{k=0}^{N=1}\left|X\left(k\right)\right|}$$ (6)

where $\left|X\left(k\right)\right|$ represents the amplitude of the k-th spectral component.

The spectral contrast measures the difference between high- and low-energy regions in the spectrum and is calculated as shown in Eqn. (7).

$$Contrast=log\left\{\frac{1}{\alpha N}\sum _{i=1}^{\alpha N}{x}_{k,i}^{\prime }\right\}-\text{log}\left\{\frac{1}{\alpha N}\sum _{i=1}^{\alpha N}{x}_{k,N-i+1}^{\prime }\right\}$$ (7)

where $\alpha$ is the neighborhood factor, which indicates the proportion of data involved in peak or trough calculations, which is set to 0.02.

The spectral flux is calculated on the basis of the squared difference between spectral amplitude vectors across frames, as shown in Eqn. (8).

$$Flux=\sqrt{\frac{1}{N}\sum _{k=0}^{N-1}{\left(\frac{x_t\left(k\right)}{E_t}-\frac{x_{t-1}\left(k\right)}{E_{t-1}}\right)}^2}$$ (8)

where $E_t$ represents the total amplitude of all the frequency bands in the current frame.

Statistical analysis of the acoustic features

To facilitate quantitative evaluation, the statistical measures were explicitly defined and reported (Afrifa et al., 2023). Each segmented vocalization sample was considered an observational unit. For each of the six acoustic features, the vocalizations recorded under heat stress and thermoneutral conditions were compared separately by age group and rearing density using Welch’s independent-samples t tests. P < 0.05 was considered to indicate statistical significance. To quantify the magnitude and practical relevance of the age-specific differences between the two conditions, Cohen’s d was calculated for each acoustic feature on the basis of the pooled sample standard deviation (Lakens, 2013). Positive Cohen’s d values indicate higher feature values under heat stress conditions, whereas negative values indicate higher values under thermoneutral conditions. The 95 % confidence intervals for Cohen’s d were estimated using 10,000 bias-corrected and accelerated bootstrap resamples. Absolute Cohen’s d values of <0.20, 0.20–0.49, 0.50–0.79, and ≥0.80 were interpreted as very small, small, moderate, and large effects, respectively.

CVC-Fusion model architecture

Overall architecture

A CVC-Fusion model based on a dual-encoder architecture is proposed, as shown in Fig. 4. The model utilizes two parallel feature extraction branches to model the 1D frequency-domain information and 2D spectrogram (time-frequency domain) information of the chicken vocalizations simultaneously. This design fully exploits the local feature extraction advantages of convolutional operators and the long-range modeling capabilities of self-attention mechanisms (Maaz et al., 2023), enabling the model to mine both the frequency-domain characteristics and the spectrogram characteristics of the chicken vocalizations. A dual-branch feature fusion module at the model’s output enables information exchange between and complementary integration of the frequency-domain and spectrogram features, facilitating the recognition of chicken vocalizations under heat stress and thermoneutral conditions.

Fig. 4.

The architecture of CVC-Fusion model.

Specifically, the two feature extraction branches in the CVC-Fusion model are the spectrogram branch and the frequency-domain branch. With respect to the spectrogram branch, referred to as the CVC-2D-HFS branch, the MobileViTv3 model (Wadekar and Chaurasia, 2022) is enhanced by replacing the MobileViTv3 block in the original model with the hybrid feature selection (HFS) block (Section 2.6.3) to improve its ability to model spectrogram features (Mehta and Rastegari, 2021). The CVC-2D-HFS branch constructs hierarchical representations of the spectrogram features by stacking multiple MobilenetV2 (MV2) blocks (Section 2.6.2) and HFS blocks. A global pooling layer is used to compress the feature map dimensions to 256 × 1, yielding a high-level representation of the spectrogram. With respect to the frequency-domain branch, referred to as the CVC-1D branch, owing to the short duration of the chicken vocalization segments, the data obtained after the signal is transformed to the frequency domain have relatively low dimensionality. To reduce the risk of overfitting and increase the effectiveness of the feature extraction process, the branch was designed as a shallow 1D convolutional feature extraction network. The process is as follows: a 1 × 3 convolution is applied to the input data of length L, generating a feature map with dimensions of 16 × L. The features are further extracted through two consecutive MV2 blocks, producing a feature map with dimensions of 32 × L/2. A 1 × 1 convolution is then used to upsample the number of channels to 64, followed by a pooling layer that compresses the feature map dimensions to 64 × 1, forming a global representation of the frequency-domain features. At the end of the CVC-Fusion model, the dual-branch feature fusion module receives the spectrogram representations from the CVC-2D-HFS branch and the frequency-domain features from the CVC-1D branch. Each of these representations is processed via a linear layer for dimensional alignment, followed by a positionwise addition operation for cross-domain fusion of the dual-branch features. Finally, the features are processed by another linear layer and a softmax function to complete the chicken vocalization classification task.

MV2 block

The MV2 block in the CVC-Fusion model adopts an inverted residual structure, as shown in Fig. 5 (Sandler et al., 2018). The MV2 block first increases the channel dimension using a 1 × 1 convolution to enhance feature expressiveness. Then, the features are extracted in a high-dimensional space using a 3 × 3 depthwise separable convolution. Finally, the features are compressed into a low-dimensional space using another 1 × 1 convolution. For the MV2 block with a stride of 1, a skip connection is employed to directly add the input and output, thereby preserving the original information and improving gradient propagation stability. For the MV2 block with a stride of 2, the output is directly produced to downsample the feature map. The MV2 block has the advantages of low computational complexity and few parameters, making it suitable for the efficient extraction of features at different levels from the chicken vocalization signal.

Fig. 5.

The architecture of MV2 block.

HFS block

The HFS block in the CVC-Fusion model is an improvement upon the MobileViTv3 block, as shown in Fig. 6. The fusion block in MobileViTv3 is replaced with a hybrid feature selection module, which dynamically weights local and global representations to further enhance feature expressiveness. The process is as follows: The HFS block receives an input feature map of size $C_{in}\times H_{in}\times W_{in}$ from the previous layer. Then, features are extracted using a 3 × 3 depthwise separable convolution and a 1 × 1 convolution, generating a local representation $F_{local}$ of size $C_{local}\times H_0\times W_0$. This local representation is fed into a linear transformer encoder, which consists of separable self-attention (SSA) mechanisms (Section 2.6.4) and a feedforward network to model long-range dependencies between features, producing a global representation $F_{global}$ of size $C_{global}\times H_0\times W_0$. Global average pooling is applied to both the local representation $F_{local}$ and the global representation $F_{global}$, resulting in feature vectors $V_{local}$ of size $C_{local}\times 1\times 1$ and $V_{global}$ of size $C_{global}\times 1\times 1$, respectively. These feature vectors are passed through a shared fully connected layer to learn the weights $W_{local}$ and $W_{global}$ for each channel, which are then normalized using the softmax function to ensure that their sum equals 1. The local and global representations $F_{local}$ and $F_{global}$ are weighted and summed using $W_{local}$ and $W_{global}$, respectively, to obtain a mixed feature $F_{mix}$ of size $C_w\times H_0\times W_0$, calculated via Eqn. (9). Finally, a 1 × 1 convolution is applied to adjust the channel dimension, generating an output feature map of size $C_{out}\times H_{out}\times W_{out}$.

$$F_{mix}=W_{local}{F}_{local}+W_{global}{F}_{global}$$ (9)

Fig. 6.

The architecture of HFS block.

SSA module

As a variant of the self-attention mechanism, the core idea of the SSA module in the HFS block is to simplify the global context modeling process by incorporating a latent token (Mehta and Rastegari, 2022), as shown in Fig. 7. By replacing the complex batch matrix multiplication and other operations in the traditional multihead self-attention mechanism with elementwise addition and multiplication operations, the SSA module offers advantages such as computational simplicity and high efficiency. The workflow is as follows: For an input feature of size k × d, where k represents the number of tokens and d is the token length, the input is processed through three branches, I, K, and V, using linear transformations. The I branch generates a latent token of size k × 1 via a linear transformation, which is then passed through a softmax function to obtain context scores of size k × 1. The K branch transforms the input into features of size k × d, while the V branch generates features of size k × d. The features from the K branch are elementwise multiplied with the context scores to generate a new feature of size k × d. This feature is then elementwise added to obtain a context vector of size 1 × d, which is subsequently elementwise multiplied with the features from the V branch to produce a new feature of size k × d. Finally, a linear transformation is applied to generate the output feature of size k × d.

Fig. 7.

The architecture of SSA module.

Model training and testing

From the original laboratory chicken vocalization dataset, 1,062 samples were selected according to the class proportions. For each of these selected samples, one of the four aforementioned data augmentation methods was randomly applied, yielding 1,062 augmented samples. These augmented samples, together with their corresponding original samples, constituted the training set, which included a total of 2,124 samples. The remaining samples in the original laboratory chicken vocalization dataset were used to construct the validation and test sets. As a result, the augmented laboratory chicken vocalization dataset contained a total of 3,033 samples, with the training, validation, and test sets split at a 7:1:2 ratio. Because the vocalizations were recorded from groups of chickens using room-level microphones, individual callers could not be identified. Therefore, the laboratory dataset was partitioned at the segmented vocalization level rather than at the individual chicken, room, or recording session level. Data augmentation was applied only to the original samples assigned to the training set; thus, no augmented counterpart of a validation or test sample was included in the training set. Similar segment- or clip-level partitioning strategies have been adopted in previous deep learning-based poultry vocalization classification studies (Mao et al., 2022; Lin et al., 2026).

During the training phase of the CVC-Fusion model, the SGD optimizer was used to update the weights and parameters of the network, while the cross-entropy loss function was employed to quantify the difference between the ground truth and the model’s predictions. The key hyperparameters that needed to be set included the number of training epochs, batch size, input size and width multiplier for the CVC-2D-HFS branch, input length for the CVC-1D branch, learning rate, momentum, and weight decay, as shown in Table 2. The source code used to train the CVC-Fusion model is publicly available at https://github.com/CoderYLH/CVC-Fusion. After the model training was completed, model performance was evaluated on the test set.

Table 2.

Parameter settings for training the CVC-Fusion model.

Parameters	Values
Training epoch	100
Batch size	12
Input size of CVC-2D-HFS	224
Width multiplier of CVC-2D-HFS	0.5
Input length of CVC-1D	100
Learning rate	0.01
Momentum	0.9
Weight decay	0.0001

Model transfer learning in a real farming environment

Owing to the significant differences in background noise, rearing density, and chicken breeds between actual chicken houses and laboratory environments, transfer learning was used to verify the applicability of the proposed CVC-Fusion model in a real-world farming environment. Specifically, the CVC-Fusion model pretrained on the augmented laboratory chicken vocalization dataset was fine-tuned using the real-world chicken vocalization dataset. In detail, the real-world chicken vocalization dataset was randomly partitioned into training and test sets at a 4:1 ratio at the segmented vocalization level. A full-parameter fine-tuning strategy was adopted: the pretrained model weights and biases from the augmented laboratory chicken vocalization dataset were loaded, and the model was trained further on the real-world chicken vocalization dataset. This approach leveraged the general latent representations of chicken vocalizations learned in the laboratory environment. Finally, the performance of the transferred model was evaluated.

Performance evaluation metrics for the CVC-Fusion model

$Accuracy$, $precision$, $recall$, and the $F_1$ score were used to evaluate the classification performance of the CVC-Fusion model (Afrifa et al., 2022; Appiahene et al., 2026).

$Accuracy$ is the proportion of all samples that are correctly classified and is calculated using Eqn. (10).

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$ (10)

$Precision$ represents the proportion of samples classified as thermoneutral vocalizations that were actually obtained under thermoneutral conditions. It therefore reflects the reliability of the model’s thermoneutral predictions and is calculated using Eqn. (11).

$$Precision=\frac{TP}{TP+FP}$$ (11)

$Recall$ represents the proportion of thermoneutral vocalization samples that were correctly identified by the model. It therefore reflects the model’s ability to avoid misclassifying thermoneutral vocalizations as heat-stress vocalizations and is calculated using Eqn. (12).

$$Recall=\frac{TP}{TP+FN}$$ (12)

The $F_1$ score is the harmonic mean of the $precision$ and $recall$ and represents the trade-off between false-positive and false-negative classifications for the thermoneutral class. It is calculated using Eqn. (13).

$$F_1=\frac{2\times Precision\times Recall}{Precision+Recall}$$ (13)

Weighted Cohen’s kappa (${\kappa }_w$) was used to measure the agreement between the predicted and true class labels after accounting for agreement expected by chance. It was calculated using Eqns. (14)–(16):

$${\kappa }_w=\frac{p_o-p_e}{1-p_e}$$ (14)

$$p_o=\frac{TP+TN}{N}$$ (15)

$$p_e=\frac{\left(TP+FN\right)\left(TP+FP\right)+\left(FP+TN\right)\left(FN+TN\right)}{N^2}$$ (16)

where $N=TP+TN+FP+FN$, $p_o$ denotes the observed agreement, and $p_e$ denotes the agreement expected by chance based on the marginal class distributions. TP denotes thermoneutral samples correctly classified as thermoneutral; FP denotes heat-stress samples incorrectly classified as thermoneutral; FN denotes thermoneutral samples incorrectly classified as heat stress; and TN denotes heat-stress samples correctly classified as heat stress.

Area under the receiver operating characteristic curve (ROC-AUC) was additionally used to evaluate the threshold-independent ability of the CVC-Fusion model to distinguish between heat stress and thermoneutral vocalizations. It was calculated from the predicted probabilities on the test set, and its value closer to 1 indicates better discriminative performance.

Results and discussion

Time-domain and frequency-domain feature analysis of chicken vocalizations

To visually analyze the differences in chicken vocalizations under heat stress and thermoneutral conditions, the time-domain signals, frequency-domain signals, and STFT spectrograms of the chicken vocalizations collected under both conditions were plotted using one representative sound sample from the experimental and control groups at a rearing density of 4 birds/m² and chickens that were 4 weeks of age, as shown in Fig. 8. To more clearly illustrate the signal characteristics, increase visual contrast, and improve interpretability, the vertical axis of the spectrograms was plotted on a logarithmic frequency scale, with the frequency range restricted to 1000–5500 Hz. The time-domain curves (Figs. 8(a) and 8(b)) reveal that the amplitude fluctuations of the chicken vocalizations under heat stress conditions were more pronounced than those under thermoneutral conditions, with a wider range of amplitude values. In the frequency-domain curves (Figs. 8(c) and 8(d)), the magnitude of the frequency-domain signals under heat stress conditions was generally greater than that under thermoneutral conditions, and the energy distribution of the signals became more concentrated within specific frequency bands under heat stress conditions. The spectrograms (Figs. 8(e) and 8(f)) revealed that the energy of the chicken vocalizations under heat stress conditions was significantly greater than that under thermoneutral conditions. The greater signal magnitude and more concentrated spectral energy under heat stress conditions may have a partly respiratory basis. High ambient temperatures induce panting and markedly increase respiratory frequency in domestic fowl (Kassim and Sykes, 1982; Brugaletta et al., 2022). In chickens, sound production at the syrinx is influenced by airflow, air-sac pressure, airway resistance, and the syringeal configuration (Brackenbury, 1980). Thus, panting-related changes in airflow and pressure could plausibly alter the sound amplitude and spectral energy distribution. Additionally, heat exposure affects vocalization-type composition and within-type acoustic characteristics in laying hens (Lin et al., 2026). However, because the respiratory rate, air-sac pressure, body temperature, and metabolic variables were not measured in the present study, this explanation remains a hypothesis that requires further physiological validation.

Fig. 8.

Time-domain signals, frequency-domain signals, and spectrograms of one representative sound sample each from the experimental group and the control group at a rearing density of 4 birds/m² and 4 weeks of age. (a) Time-domain signal of the sound sample in the experimental group; (b) Time-domain signal of the sound sample in the control group; (c) Frequency-domain signal of the sound sample in the experimental group; (d) Frequency-domain signal of the sound sample in the control group; (e) Spectrogram of the sound sample in the experimental group; (f) Spectrogram of the sound sample in the control group.

In this study, the time-domain and frequency-domain features of the chicken vocalizations under heat stress and thermoneutral conditions were calculated for different age groups, as shown in Fig. 9. The corresponding Cohen’s d values and 95 % confidence intervals are presented in Table S1. The STE (Fig. 9(a)) of the chicken vocalizations under heat stress conditions was significantly greater than that under thermoneutral conditions, with large positive effects across all age groups (Cohen’s d = 1.469–1.812), indicating that the overall energy level of the vocalizations was greater under heat stress conditions. This may be related to changes in metabolism and the respiratory rate caused by heat stress (Brugaletta et al., 2022). Additionally, the ZCR (Fig. 9(b)) of the vocalizations from different age groups was greater under thermoneutral conditions than under heat stress conditions, suggesting that the number of zero-crossings was reduced in the vocalizations collected under heat stress conditions. Furthermore, the spectral centroid (Fig. 9(c)) of the chicken vocalizations was greater under thermoneutral conditions than under heat stress conditions, indicating that the frequency distribution of the vocalizations tended to shift toward higher frequency bands under thermoneutral conditions than under heat stress conditions. Both the ZCR and the spectral centroid showed large negative effects across all age groups (Cohen’s d = −1.989 to −0.981 and −1.977 to −0.896, respectively). Moreover, the spectral flatness (Fig. 9(d)) was lower under heat stress conditions than under thermoneutral conditions, with a moderate negative effect at an age of 4 weeks and large negative effects at ages of 5 to 8 weeks, suggesting a more concentrated frequency distribution under heat stress conditions. Except at 5 weeks of age, the spectral contrast of the chicken vocalizations in the experimental group was greater than that in the control group (Fig. 9(e)), with large positive effects at ages of 4, 6, 7, and 8 weeks, indicating that the differences in energy across different frequency bands under heat stress conditions were greater than those under thermoneutral conditions. In addition, spectral flux was significantly greater under heat stress conditions at 4, 6, 7, and 8 weeks of age, although the effect was small at 4 weeks of age and moderate from 6 to 8 weeks of age. No significant difference in spectral flux was observed at 5 weeks of age (Fig. 9(f)). The relatively weak effects of heat stress on spectral flux at 4 and 5 weeks of age, together with the very small effect of heat stress on spectral contrast at 5 weeks of age, may be influenced by age-related variations in baseline vocalization patterns. Previous studies have shown that broiler vocalization frequency and spectral patterns change with age and body weight (Fontana et al., 2016; Carvalho et al., 2025). However, because individual growth variables and vocal-organ development were not measured in the present study, a specific growth-related mechanism cannot be confirmed and remains a hypothesis.

Fig. 9.

Comparison of time-domain and frequency-domain features of chicken vocalizations under heat stress (experimental group) and thermoneutral conditions (control group) at different age groups. (a) STE in the time domain; (b) ZCR in the time domain; (c) Spectral centroid in the frequency domain; (d) Spectral flatness in the frequency domain; (e) Spectral contrast in the frequency domain; (f) Spectral flux in the frequency domain.

The differences in the time-domain and frequency-domain features of the chicken vocalizations under heat stress and thermoneutral conditions indicated that heat stress not only alters the energy and frequency distributions of the vocalizations but also affects their dynamic characteristics. These findings provide important evidence for the development of a vocalization recognition model for chickens under heat stress conditions.

Furthermore, the time-domain and frequency-domain features of the chicken vocalizations under heat stress and thermoneutral conditions were calculated at different rearing densities, as shown in Table 3. Compared with that under thermoneutral conditions, the STE of the chicken vocalizations increased significantly under heat stress conditions, with more pronounced changes observed at higher rearing densities. Additionally, compared with those under thermoneutral conditions, the ZCR decreased, the spectral centroid shifted toward lower frequency bands, the spectral flatness decreased, the spectral contrast increased, and the spectral flux increased under heat stress conditions. These results indicated that heat stress significantly affected chicken vocalizations at different rearing densities, as reflected by increased energy, a more concentrated spectral distribution, and increased changes in dynamic characteristics, which was consistent with the effects of heat stress observed in different age groups.

Table 3.

Comparison of mean time-domain and frequency-domain features of chicken vocalizations under heat stress (experimental group) and thermoneutral conditions (control group) at different rearing densities.

Groups	Rearing density (birds/m²)	STE ($\times {10}^{-4}$)	ZCR ($\times {10}^{-1}$)	Spectral centroid ($\times {10}^3$)	Spectral flatness ($\times {10}^{-1}$)	Spectral contrast ($\times 10$)	Spectral flux ($\times {10}^0$)
Experimental	4	7.70***	3.90***	10.01***	3.50***	1.30***	1.07***
control	4	5.29	5.02	11.70	4.09	1.22	1.04
Experimental	8	15.00***	3.66***	9.82***	2.98***	1.28***	1.07
control	8	5.27	5.02	11.74	4.08	1.23	1.06
Experimental	12	13.36***	3.60***	9.75***	3.06***	1.30***	1.10***
control	12	4.75	4.86	11.50	4.17	1.23	1.06

Note: ***p < 0.001, **p < 0.01, *p < 0.05.

Performance analysis of the CVC-Fusion model

Performance of the CVC-Fusion model

Fig. 10 shows the confusion matrix of the CVC-Fusion model on the augmented laboratory chicken vocalization dataset. The model achieved an $accuracy$ of 96.86 %, a $precision$ of 97.03 %, a $recall$ of 96.71 %, an $F_1$ score of 96.87 %, a weighted Cohen’s kappa (${\kappa }_w$) of 0.937, and an ROC-AUC of 0.986, indicating strong classification performance, strong agreement beyond chance, and excellent discriminative ability. The confusion matrix reveals that the model’s performance was balanced for both classes. The number of vocal samples correctly identified by the model under heat stress conditions was very close to the number correctly identified under thermoneutral conditions. However, the model mistakenly classified 10 vocal samples under thermoneutral conditions as heat stress samples and 9 vocal samples under heat stress conditions as thermoneutral samples.

Fig. 10.

Confusion matrix of the CVC-Fusion model on the augmented laboratory chicken vocalization dataset.

To further investigate the causes of misclassification, 19 misclassified samples from the dataset were separately extracted, and time-domain and frequency-domain feature analyses were performed as described in the Methods for calculating audio time-domain and frequency-domain features section. These misclassified samples exhibited certain similarities in terms of both the time-domain and frequency-domain features. Specifically, the amplitude fluctuations of the 10 misclassified thermoneutral samples were similar to those of the heat stress samples in the time domain, and the energy distribution was more concentrated in the frequency domain. Additionally, the 9 misclassified heat stress samples exhibited more stable signal characteristics in the time domain, with a more dispersed energy distribution in the frequency domain, resembling the characteristics of the thermoneutral samples. These feature similarities reduced the model’s ability to distinguish between the two types of vocalizations in the high-dimensional feature space, leading to misclassification. The biological origins of these boundary samples cannot be determined from the acoustic data alone. Possible explanations include temporary recovery or adjustment following heat exposure and individual variations in heat tolerance. However, because individual callers were not identified and their physiological states were not measured, these interpretations require further biological validation. In future studies, individual-level vocalization tracking should be combined with synchronized physiological and behavioral measurements across heat exposure and recovery periods. Recent digital agri-food research has similarly emphasized the importance of multimodal integration and real-world validation for assessing model robustness and generalizability (Afrifa et al., 2026).

Further analysis suggested that the misclassification could be attributed to two factors: First, the lower distinguishability of the features between the two classes in both the time domain and the frequency domain, which made it difficult for the feature extraction module to capture discriminative representations; second, in the high-dimensional feature space, the feature distributions of these samples may overlap to some extent, making accurate separation of the two classes at the decision boundary by the classifier challenging. To improve the model’s ability to recognize such confusing samples, feature enhancement mechanisms should be integrated in future work, such as increasing the weights of the boundary samples in the training data or using adversarial training methods to improve the model’s discriminative ability.

Ablation study of the CVC-Fusion model

To investigate the contributions of different modules in the CVC-Fusion model, ablation experiments were performed by progressively removing or replacing key components of the model to assess their impact on overall performance, as shown in Table 4. On the basis of Experiments 1 and 3, when the single-branch CVC-2D model or CVC-1D model was used for chicken vocalization recognition, the $accuracy$ values were 88.12 % and 94.39 %, respectively. The $accuracy$ of the CVC-1D model was significantly greater than that of the CVC-2D model, indicating that frequency-domain features were more easily captured and classified by the model than spectrogram features were. Although the spectrograms contained richer information, they also introduced additional noise, which interfered with the model’s decision-making ability. In contrast, the frequency-domain features provided more compact and focused representations, allowing the model to more effectively recognize different chicken vocalizations. On the basis of Experiments 1 and 2, when the HFS module was integrated into the CVC-2D model, the $accuracy$ increased from 88.12 % to 89.93 %. This result suggested that the HFS module enhanced the model’s ability to model spectrogram features by dynamically weighting local and global features, thereby improving classification performance. On the basis of Experiments 3 and 4, when the CVC-2D branch and CVC-1D branch were fused, the $accuracy$ increased by 1.81 % compared with that of the single-branch CVC-1D model. This finding indicated that the dual-branch architecture effectively integrated feature information from both the frequency domain and spectrograms. In Experiment 5, when the CVC-2D, HFS, and CVC-1D components were combined in the complete CVC-Fusion model, the $accuracy$ increased further to 96.86 %. Although this represented a modest improvement of only 0.66 % compared with the accuracy obtained in Experiment 4, it highlights that the HFS module, with its dynamic weighting mechanism for spectrogram features, further optimized the model’s decision-making ability and increased its robustness when handling boundary samples.

Table 4.

Contribution of different modules to the CVC-Fusion model.

Experiment	CVC-2D	HFS	CVC-1D	CVC-Fusion	$Accuracy$/%
1	√	×	×	×	8 8.12
2	√	√	×	×	89.93
3	×	×	√	×	9 4.39
4	√	×	√	√	96.20
5	√	√	√	√	96.86

These results demonstrated that the CVC-Fusion model achieved optimal performance when the strengths of the CVC-2D, CVC-1D, and HFS components were integrated, thereby validating the effectiveness of the model architecture.

Comparative experiments of the CVC-Fusion model

To comprehensively evaluate the performance of the CVC-Fusion model, the model’s superiority and the rationality of its module designs were comprehensively analyzed through comparisons of the use of different attention mechanisms within the CVC-Fusion model, the dual-branch feature fusion approach, and its performance relative to that of other models. The effects of different attention mechanisms within the CVC-2D-HFS branch of the CVC-Fusion model on model performance are presented in Table 5. The model using the traditional self-attention mechanism achieved an $accuracy$ of 89.77 % (Dosovitskiy et al., 2020), whereas the model employing Linformer attention achieved a slightly lower $accuracy$ of 88.28 %. This difference may be due to the use of low-rank approximations in the Linformer attention mechanism when modeling global features, which reduced its ability to capture detailed features (Wang et al., 2020). In contrast, the model using the SSA mechanism achieved the highest $accuracy$ of 89.93 %. Among the three attention mechanisms, the SSA mechanism performed the best. This indicated that the SSA mechanism, while maintaining relatively low computational complexity, was effective in modeling the global dependencies between features.

Table 5.

Impact of attention mechanisms in the CVC-2D-HFS branch of the CVC-Fusion model on model performance.

Attention type	$Accuracy$/%
Self-attention	89.77
Linformer attention	88.28
SSA	89.93

Additionally, the performance of CVC-Fusion models using different dual-branch feature fusion methods was compared, as shown in Table 6. The model achieved an $accuracy$ of 96.86 % when the dual-branch features were directly added, a simple fusion method that eliminated the need for additional parameters. The model using feature concatenation for dual-branch fusion achieved an $accuracy$ of 96.37 %, which was slightly lower than that of the model using the simple addition method. Additionally, the concatenation operation increased the feature dimensions, resulting in increased computational complexity. The performance differences of the models with the unidirectional attention mechanisms indicated that features from different branches played distinct roles in the feature fusion process. Compared with the model using SSA_q2d, the model using SSA_q1d had a significantly lower $accuracy$, which suggested that frequency-domain features were more suitable for serving as keys and values in the attention calculation. When spectrogram features are used as queries, the global information of the frequency-domain features can be flexibly selected, thereby improving the fusion performance. The model using the bidirectional attention mechanism (Bidirectional SSA) achieved the highest $accuracy$ of 97.03 %, as it further leveraged the complementarity between the features of the two branches. However, compared with the simple addition method, the accuracy increased by only 0.17 %, while the computational complexity and number of parameters increased significantly. Therefore, the simple addition approach was adopted as the final fusion strategy in the CVC-Fusion model, as this approach effectively balanced performance and complexity.

Table 6.

Impact of dual-branch feature fusion methods in the CVC-Fusion model on model performance.

Fusion method	$Accuracy$/%
Sum	96.86
Concat	96.37
SSAq1d	90.26
SSAq2d	96.04
Bidirectional SSA	97.03

Note: SSA_q1d indicated that the CVC-1D branch features were used as the query, and the CVC-2D-HFS branch features were used as the key and value for the SSA computation. SSA_q2d indicated that the CVC-2D-HFS branch features were used as the query, and the CVC-1D branch features were used as the key and value for the SSA computation. Bidirectional SSA indicated that the CVC-1D branch features and CVC-2D-HFS branch features were used as mutual queries, keys, and values for the SSA computation.

Furthermore, the performance of the CVC-Fusion model was compared with that of other lightweight models in the chicken vocalization recognition task under heat stress and thermoneutral conditions, as shown in Table 7. Overall, the CVC-Fusion model outperformed the comparison models in terms of all four metrics, achieving $accuracy$, $precision$, $recall$, and $F_1$ score values of 96.86 %, 97.03 %, 96.71 %, and 96.87 %, respectively, demonstrating its superior overall performance. Among the 2D-only models, the CVC-2D-HFS model achieved an $accuracy$ of 89.93 %, significantly outperforming ResNet-18, EfficientNet v2, and ShuffleNet v2 (He et al., 2015; Ma et al., 2018; Tan and Le, 2021). With respect to the 1D-only models, the CVC-1D model achieved an $accuracy$ of 94.39 %, outperforming the other 1D-only models while also having significantly fewer parameters than the other 1D-only models, demonstrating its effectiveness and efficiency in terms of feature extraction. Furthermore, the performance of the CVC-Fusion model was further improved by integrating features from both the CVC-2D-HFS and CVC-1D branches, indicating that the information complementarity mechanism of the dual-branch structure played a key role in the chicken vocalization classification task. The CVC-Fusion model required only 0.88 M parameters, which was significantly lower than the number of parameters required by the deep learning models such as ResNet-18, emphasizing the efficiency and lightweight design of the proposed approach. These results validated that the CVC-Fusion model achieved high performance while maintaining remarkable computational efficiency, making it suitable for resource-constrained scenarios.

Table 7.

Performance comparison of different chicken vocalization recognition models.

Model type	Model	$Accuracy$/%	$Precision$/%	$Recall$/%	$F_1$/%	Params/M
2D-only	Resnet-18	87.13	87.46	86.87	87.16	11.18
	Efficientnet v2	88.28	88.82	87.93	88.37	5.86
	ShuffleNet v2	87.79	87.34	88.52	87.93	0.34
	CVC-2D-HFS	89.93	89.52	91.00	90.25	0.79
1D-only	Resnet-18	94.22	93.81	95.05	94.43	0.06
	Efficientnet v2	93.89	93.65	93.96	93.80	1.63
	ShuffleNet v2	93.07	92.67	93.29	92.98	1.24
	CVC-1D	94.39	94.00	94.63	94.31	0.01
Fusion	CVC-Fusion	96.86	97.03	96.71	96.87	0.88

Visualization of features in the CVC-Fusion model

To elucidate the decision-making mechanism of the CVC-Fusion model for chicken vocalization recognition under heat stress and thermoneutral conditions, the Grad-CAM technique was applied to visualize the feature extraction process of the dual-branch components of the model (Selvaraju et al., 2017), as shown in Fig. 11. Images of the experimental and control group samples after preprocessing, which were then fed into the CVC-2D-HFS branch of the CVC-Fusion model, are shown in Figs. 11(a) and 11(b), respectively. The heatmaps generated by visualizing the features output by the final feature extraction module of the CVC-2D-HFS branch for the experimental and control group samples are shown in Figs. 11(c) and 11(d), respectively. The high-attention red regions in the heatmaps closely matched the high-energy regions in the original spectrograms, indicating that the CVC-2D-HFS branch effectively captured the energy features in the key frequency bands of chicken vocalizations, particularly in areas with concentrated energy. The weight curves generated by visualizing the features output by the final feature extraction module of the CVC-1D branch for the experimental and control group samples are shown in Figs. 11(e) and 11(f), respectively. The curves represent the frequency domain of the original samples, with solid points indicating the weights of the frequency-domain data. As shown in Fig. 11(e), the first three data points at the left end of the curve were assigned relatively high weight values, indicating a potential focus on specific features. In contrast, the weights of the data points in the remainder of the curve were distributed more evenly, reflecting the focus of the CVC-1D branch on the global frequency-domain features of chicken vocalizations under heat stress conditions. As shown in Fig. 11(f), the CVC-1D branch assigned higher weights to the data points at the two ends of the curve (frames 4–12 and frames 27–41) while paying relatively less attention to the middle peak region (frames 13–26). Notably, the middle peak region corresponded to the energy-concentrated regions in the spectrogram, indicating that for chicken vocalizations under thermoneutral conditions, the features extracted by the CVC-2D-HFS and CVC-1D branches provided complementary information: the CVC-2D-HFS branch focused on the energy-concentrated areas in the frequency domain, whereas the CVC-1D branch focused on areas with relatively lower magnitudes.

Fig. 11.

Visualization of dual-branch weights of the CVC-Fusion model under heat stress (experimental group) and thermoneutral conditions (control group). (a) Spectrogram of the sound sample in the experimental group; (b) Spectrogram of the sound sample in the control group; (c) Heatmap of the sound sample in the experimental group; (d) Heatmap of the sound sample in the control group; (e) Magnitude curve of the sound sample in the experimental group; (f) Magnitude curve of the sound sample in the control group.

These visualization results demonstrated that the CVC-Fusion model can effectively extract key features from both the spectrogram and frequency-domain curves through the synergistic effect of the two branches, providing reliable decision-making support for chicken vocalization recognition tasks. These results further validated the rationality of the model design and its adaptability to input signals, highlighting the interpretability and application potential of the CVC-Fusion model.

Results of model transfer learning in a real-world farming environment

The CVC-Fusion model achieved an $accuracy$ of 80.69 %, a $precision$ of 80.70 %, a $recall$ of 80.69 %, an $F_1$ score of 80.62 %, a weighted Cohen’s kappa (${\kappa }_w$) of 0.607, and an ROC-AUC value of 0.794 on the real-world chicken vocalization dataset. Compared with its excellent performance in the controlled laboratory environment, the model faced challenges in accurately identifying heat-stress vocalizations under real-world farming conditions. This was primarily manifested as cross-misclassification errors between vocalizations recorded when most of the chickens exhibited panting behavior under heat stress conditions and those recorded when most chickens displayed normal behavior without typical stress responses. These results indicate that the feature distributions of the two classes exhibit significant overlap in boundary regions in the high-dimensional feature space.

This phenomenon may be attributed to two main factors. First, compared with the strictly controlled environmental parameters in the laboratory setting, the environmental temperature and humidity in the semiopen chicken house are affected by multiple factors, such as time, weather, and the local rearing density. Consequently, the thermoregulatory mechanisms underlying the effects of heat stress on the vocalization patterns and features of chickens are dynamic; the sounds produced under mild heat stress conditions are acoustically similar to those produced under thermoneutral conditions. Second, the sound sources and acoustic components in the semiopen cage-free chicken house are highly complex. Even with the assistance of synchronized RGB video, the lack of a fine-grained sound source localization system inevitably introduced subjective judgment errors during the manual labeling process, thereby introducing label noise.

These findings demonstrate that achieving precise heat-stress recognition in real-world farming environments using a single audio modality remains challenging. Therefore, in future research, additional modalities (e.g., RGB and infrared) should be integrated to develop more accurate and robust chicken heat-stress recognition methods. Moreover, because both datasets used in the study were partitioned at the segmented vocalization level, the reported results reflect sample-level within-dataset performance. Therefore, in future evaluations, nonoverlapping recording sessions and independent flocks, chicken houses, and farms should be used.

Conclusions

In this study, compared with those recorded under thermoneutral conditions, chicken vocalizations recorded under heat stress conditions generally presented greater energy and more concentrated spectral distributions. The lightweight CVC-Fusion model effectively recognized chicken vocalizations under heat stress and thermoneutral conditions in the laboratory environment, achieving an $accuracy$ of 96.86 %, a weighted Cohen’s kappa of 0.937, and an ROC-AUC of 0.986 with only 0.88 million parameters. Grad-CAM visualization revealed that the dual-encoder architecture captured complementary information from frequency-domain features and spectrograms, supporting the effectiveness of the model design. After transfer learning, the CVC-Fusion model achieved an $accuracy$ of 80.69 %, a weighted Cohen’s kappa of 0.607, and an ROC-AUC of 0.794 in a real-world environment. This lower performance was likely attributable to the complex sound sources and acoustic components in the semiopen cage-free chicken house and highlights the limitations and challenges of heat-stress recognition based on a single audio modality. In future research, audio data could be combined with RGB or infrared information, and biologically independent evaluations could be conducted using nonoverlapping recording sessions and independent flocks, chicken houses, and farms.

Disclosures

The authors declare no conflicts of interest.

Author contribution

Yilei Hu made substantial contributions to the experimental investigation and data acquisition, methodological development, software implementation, and preparation of the original manuscript draft. Jinyang Xu made substantial contributions to software implementation, model validation, and the analysis and interpretation of the results. Jinming Pan made substantial contributions to the methodological design of the study. Di Cui conceived and supervised the study, secured funding, contributed to the interpretation of the results, and critically revised the manuscript for important intellectual content. All authors critically reviewed the manuscript for important intellectual content, explicitly approved the exact final version for publication, and agree to be personally and publicly accountable for all aspects of the work, including ensuring that any questions concerning the accuracy or integrity of the work are appropriately investigated and resolved.