Raven Calls Indicate Sender’s Neural State

Vocal communication accounts for dominantly percentage within animal species. The information of vocal samples contains not only the amplitude of objects, but also the emotional states behind it. However, to extract the emotion state behind the sound remains controversial. Here we introduce an artificial network method, the Back Propagation Neural Network, BPNN, to classify the emotional states behind the sound. The results disclosed the behaviour categories, including alarm, flight, begging and singing which has been successfully classified. This artificial intelligence classification may aid us to distinguish the ecological categories via animal vocal communication and to discover its significance of evolution and nature.


Background
The sound is one of universal phenomenon, which does not only exist in numerous nature phenomenon [1][2][3], but also appear in animal communication in species or across species [4,5]. From the perspective of physics, the sound is generated by the amplitude of object, which determines the frequency and intensity [6][7][8]. According to evolution, the ability of adaption and survival skills of species had been constructed and strengthened for surviving, for example: track the life source, avoid the harmful events and defend the territory [9,10].
With respecting the animal communication, there are three domain categories, sound, behaviour and the order [11][12][13][14]. The pronunciation process of animal depends on a series of muscles interaction, including breathing muscles, mouth muscles and throat muscles and the amplitude of vocal cord [15,16]. Vocal communication accounts for dominant percentage in species [17,18]. Sound response on the stimulation is the best indicator of sender's emotional state under the specific circumstances, such as nervousness, sadness, happiness, repulsiveness, angriness and afraid [19,20].
Interestingly, the raven, an intelligent avian, has been used in many specific mind and brain studies for decades [21][22][23]. For example, the raven's memory predict future [24], sequence tool used to get the award [25], the awareness of density of article [26], and so on. Ravens' call often appears frequently during their foraging [27], defend the territory [28], rutting and mating [29], infant parenting [30]. Even though we can hear various different callings, however, to classify the emotional states behind the sound remains unclear. Here are four behaviour validated related calls of raven are hypothesized to be directly different among acoustic parameters, providing cues to emotional category recognition. Therefore, we use the method, termed as "Back Propagation Neural Network, BPNN", to disclose the emotion behind the sound. This may be helpful to us to understand each other better, not only in same species, but also across species.

Data Source
Total 851 sound samples of raven have been downloaded from the website

Acoustic Parameters Extraction
Here we use the MATLAB script to extract the acoustic parameters [31,32].

Data Analysis
According to the raven's call recorded under the different circumstances, the biological meanings may be presented differently. Therefore, we selected Back Propagation Neural Network (BPNN) method [33][34][35]. A typical BPNN consists of the input layer, hidden layer and output layer ( Figure 1).

Result
Totally we have selected four different categories call of raven, including alarm, begging, fighting and singing. The descriptive acoustic parameters extracted from MATLAB, please see Table 1. Secondly, pre-processing: Due to realistic complexity, we used the standard process of deviation normalization (x. std. = (x-min) / (max-min)) to perform the better neural via dimensionless processing. Thirdly, primary parameter selection: The number of input layer units was 14. There was only one output layer neuron. Then, the number of hidden layer neurons was 12. Details please see Figure 2.

Figure 2. Neural Network Pattern
The unique purpose of BPNN construction is to classify. Double transfer sigmoid function and linear were used. The results were presented well when used "tansig" function in the input layer-hidden layer and linear function in the hidden layer-output layer. Furthermore, the "trainlm" method (combined gradient descent method with Newton method) was used for training, which terminated when achieved the best effect. The details of training performance process please see Figure 3.  inferred that the prediction of the model is more accurate, while it has a certain reference value. The results disclosed that the average absolute error and the average relative error of the actual classification and the predicted classification were 1.0e-09 and 1.0e-10, respectively.

Discussion
The sound is the best indicator to reflect the property of the object [36,37]. From the physical perspective, the size of sound depends on the frequency of object, and the pitch of sound depends on the amplitude of object [38,39].
With respect to the animal communication, the pronunciation is not only depending on a series of pronounced organ coordination work [40][41][42], but also depending on the control of central nervous system [43]. The animals have been adapted to the environment after a long time evolution process via a series of successful constructed conditional reflexes [44]. It is obvious that the emotional information maybe hide behind the vocal communication.
The raven is recognized as the most intelligent avian [45,46]. It was being used in mind, sequenced tool used test, awareness, learning and memory researches. More interestingly, Markus et al distinguished the gender and age of raven [47], and reported that ravens can predict future via their memory [24]. There are four behaviour categories of raven's call, including alarm, begging, flighting and singing. It is necessary to distinguish the specific behaviouristic category. Thanks to BPNN method, their behaviour categories have been classified successfully. The abundance information is contained within the sound in animal communication. There is not only the amplitude information of object, but also embedded emotional state. For example, professional police can distinguish the criminal suspects via sophisticated conservation due to their unstable irregular speaking [48]; experienced hunter can determine the degree of hunger of animal [49].
To distinguish the emotional state behind the sound is of great significance.
Through the emotion information behind the ravens' sound, we can get better understanding of their biological meanings. For instance, the alarm call often occurs on the occasion of dangerous events approaching, it is good for avoiding the harmful events and surviving [50,51]. The flight call occurs on the situation of fighting process; the failure subject presents submission to dominant subject to avoid the further damage on its body or relationship [52,53]. Moreover, the begging call happens when sub-dominant subject paying / begging for the life sources, such as the food, resting place etc [54,55]. As for the singing; this may arise on the situation of pleasant animal communication, as for the specific motivation remains controversial.
Taken together, this artificial network could be helpful of trying to distinguish the behaviour categories of different raven's sounds. It can strengthen our understanding on their biological and evolution meanings.

Conclusion
Taken together, we herewith disclose the method of BPNN could be the promising candidate to classify the category of raven's call. Furthermore, this may help us to investigate more in-depth of neurobiology of raven's mind.

Abbreviations List:
BPNN: Back Propagation Neural Network

Declarations Ethic Statement
Not applicable.

Consent for publication
Not applicable

Availability of Data and material
All data will be open for reasonable request.

Competing interests
The authors declare that they have no competing interests.
Authors' Contribution: X.F drafted the general idea and drafted the manuscript with HZX; XJ, ZJ, YCY and DXL collected the sound data from xeno-canto website and data clean.HZX and WZL performed the MATLAB analysis.