The command of comfort in an intelligent building by speech classification and image classification for energy optimization

It is very difficult to define comfort in the context of building standards of intelligent buildings. Approaches that define thermal comfort give two definitions of comfort. The first defined by Fanger (2009), who defines comfort as a state of neutrality, means a state of mind expressing the satisfaction of its environment, in this case the person cannot say if he wants to increase heating (having warmer) or air conditioning (getting colder). The second, defined by Givoni and Izard (1978) defines comfort as the conditions for which the self-regulating mechanisms of the body are at a minimum level of activity. The intelligent building has a very important role, including measuring comfort. Saizmaa and Kim (2008) analyze the issue of smart building design. They developed the notion that associates the intelligent building with psychological, emotional, physical, social, and spiritual appropriations. Ekambi Schmidt (1972) was interested in the relationship between smart buildings and the people who occupy them, and the representation that these people construct of their place of life. Moser (2009) describes this relationship as the basis in home, a physical and abstract space in which the intelligent building must be able to express and identify itself. There are a lot of projects devoted to the use of speech-recognition technologies, such as Companions (Cavazza et al., 2010), Companion Able (Rougui et al., 2009), Aladin (Gemmeke et al., 2013), or Pipin (Casanueva et al., 2014). However, there are still many questions about the real impact both in terms of assistance and installation cost, despite real evaluations in test apartments with Pers projects (Hamill et al., 2009), Dirha (Ravanelli and Omologo, 2014). Our contribution in this field of research is the realization of a robust and optimum system for the control of comfort in an intelligent building. The speech classification system is a very robust system that gives very good results for the control of comfort in an intelligent building. Its advantage over other existing solutions is that it gives the comfort desired by the person at all times and with great precision, for example, if we compare it with another system that uses the behavior of people by camera (Ahmed et al., 2013), to give them the desired comfort, the latter is not as precise as the speech classification system, because a behavior does not reflect 100% the desired comfort, but the person’s speech that reflects 100% the desired comfort is the reason that prompted us to use the classification of speech to command comfort. But after using this system, we noticed that it has some faults, such as in some situations, it will not be able to command the comfort, and these situations are:

If the person is sick to a point where he will not be able to speak, or if his voice has changed a lot, then in this case, the system will not be able to command the comfort and the energy will not be optimized.

The solution to these defects is to use the speech classification system with an image classification system, to add another image classification system in parallel with the speech classification system, and use these two systems at the same time for comfort command.

This system of classification of images will allow to solve the defects of the system of classification of speech. We have noticed that to solve these defects, the image classification system must classify three images:

The first image expresses the presence of the person.

When this system detects the presence of a person, then the intelligent building comfort management system gives optimum comfort to the person. We need this system in the case of a sick person who cannot speak so that he cannot command the comfort. If the system does not detect the presence, the intelligent building comfort management system turns off all lights and other equipment in that room, and lowers the heating or the air conditioner.

In the event that the system detects that the person is asleep, then the intelligent building comfort management system turns off all lights and other equipment in that room. The disadvantage of the image classification system is that it requires lighting, and therefore in the dark, the person will not be able to have his comfort; in addition, there are people who prefer to stay in the dark and to have their comfort. The solution to this inconvenience is the speech classification system, and therefore we notice that these two systems complement each other.

In the first system of speech classification, an experiment will be implemented using an input database, the vocabulary used in this database is made up of 20 words in the English language. The classification will be of global type, the voice signal will be represented using the MFCC (Mel Frequency Cepstral Coefficients) coefficients to form the input vector of the classification system, and the SVMs will be used for the learning and the voice recognition phase. In the second image classification system, the SIFT (Scale Invariant Feature Transform) method is used for the preparameterization of the images, and SVM will be used for the learning and the image recognition phase. We opted for SVM (Support Vector Machines) because of their good generalization capacities in many problems, but also for their good results on sets of small training data (unlike neural networks that require a larger database). They have shown in recent years their power, especially their computing time, which is very satisfactory compared to other techniques (Joachims, 1999).

In order to improve performance, a study on several nuclei (Linear, RBF, Polynomial, and Sigmoid) was carried out in the learning phase, taking the nucleus, which gives the best results.

SVM is a classification method which was introduced by V. Vapnik (1995). Support vector machines are known as support vector networks, which is a supervised learning used in machine learning (Srinivasan and Rajakumar, 2017). SVM is a classification technique based on kernel methods that has been proved very effective in solving complex classification problems in many different application domains (Dixon and Candade, 2008). This method is characterized by solid theoretical foundations which make its success. There is indeed a direct link between the theory of statistical learning and the SVM learning algorithm. The drawback of most ML (Machine Learning) techniques is that they have a large number of learning parameters to be set by the user (structure of a neural network, gradient update coefficient, etc.). In addition, with these methods, the number of parameters to be calculated by the learning algorithm is linear, even exponential, with the dimension of the input space.

An SVM teaches a separator. This brings the problem back to the definition of a separator. Let us give a finite set of vectors of Rn, separated into two groups, that is to say into two classes. Membership in one class or another is defined by a label, associated with each of the vectors, on which is written ‘class 1’ or ‘class 2’. Finding a separator is like building a function, which takes a vector from our set, and can tell which class it is. SVMs are a solution to this problem, as would be a simple rote learning of the classes associated with the vectors of our set.

For the learning, the principle of SVM, based on the theorem of statistical learning, consists first of all in transforming the inputs into a space, thanks to the kernel, we take a linear decision border in this space (Karush, 1939). Support vector machines are known as Support vector networks, which is a supervised learning used in the learning (Srinivasan and Rajakumar, 2017). Among the SVM models, there are linearly separable cases and non-linearly separable cases. The implementation of SVM classifiers in the context of the linearly separable case is simple. In most real problems, there is no possible linear separation between data, the classifier (hyperplane separator) cannot be used because it only works if the classes of training data are linearly separable. To overcome the drawbacks of non-linearly separable cases, the idea is to use the kernel function which consists in changing the data space, by performing a transformation of the representation space of the input data into one more space. Large dimension (possibly of infinite dimension), in which these input data are linearly separable. In practice, a few families of configurable kernel functions are known, and it is up to the SVM user to carry out tests to determine which one is best suited for his application. We can cite the following examples of nuclei: polynomial, Gaussian, sigmoid, and RBF (radial basic function).

In most real problems, there is no possible linear separation between the data of the input data base, which is the case for our data base. It is for this reason that we use the kernel functions so that our data is separable. For the implementation, we use several kernels in order to take the best models which give us the highest classification rate.

For the recognition SVM (Support Vector Machine) aims to find the two best hyperlinks for separating word or image into two classes. To make a classification you must: determine an appropriate classification system, select training samples, image pre-processing, extract features, select fitting classification approaches, post-classification processing and accuracy assessment (Prasad et al., 2017). The classification of a word or image is given by its position in relation to the best hyperplane (model or classifier) found in the learning phase. We used the so-called ‘one against all’ approach for classification, which successively compares one class to all the others using a binary classifier.

We conducted experiments in the Domus intelligent building at the LIG laboratory, located at the University of Grenoble, France, specifically at the CTL (Software Development Center).

Our objective is to set up a complete video surveillance network that includes two parts: the first part, consists in fixing two cameras in each room of the intelligent building, these are connected to a computer for the acquisition of permanent video sequences. The second part consists in recording the video sequences for 30 seconds, then converting them into images. Figure 10 represents the three images to be classified by the image classification system, I₁ : the presence of the person, I₂: a sleeping person, I₃: the absence of the person.

Figure 10:

Recognition

As we have three images to recognize for the image classification system, then the SVMs will generate three models (classifiers or hyperplane separators) in the learning phase, and each model corresponds to an image. Since we have three classes, then we will have three binary classifiers, which are constructed (three hyperplanes, or three problems of optimization), the third classifier being intended to distinguish the class of index 3. Each image to be classified will be presented to the three classifiers and is assigned the label of the classifier, which returned the highest percentage of adhesion. The diagram of the image classification system by SIFT and SVM is shown in Figure 11.

Figure 11:

Resulte discussion

(4) The rate of recognition / I i = NRI i NTI i , NRI_i: represents the number of recognized image I_i.N_TI_i: represents the total number of the image I_i in the base of tets. RR: represents the rate of recognition.

The precision for a I_i image is the number of images correctly detected in divided by the total number of images in the entire test base that are detected and assigned to the image I_i: (https://fr.wikipedia.org/wiki/Pr%C3%A9precision_et_rappel).

NcdIi: the number of images correctly detected for the image I_i.

NdIi: the number of images detected for the image I_i in the test base.

So: (5) Precision / I i = N cd I i N d I k I i .

The reminder for a I_i image, this is the number of correctly detected images for the I_i image divided by the total number of detected t images that belong to the I_i image (https://fr.wikipedia.org/wiki/Pr%C3%A9precision_et_rappel).

N_cd I_i: the number of images correctly detected for the image I_i.

N_T I_i: the total number of detected images that belong to image I_i.

So: (6) Reminder / I i = N cd I i NTI i . (6)

As long as the values of the recognition rate, precision, and recall are high, we can guarantee its reliability and accuracy.

The classification is done by the SVM. To improve this classification, we used 4 nuclei, linear, polynomial, RBF (radial basic function), and sigmoid.

The results of the experiments are given in Figures 12-15 and Tables 5-8.

Table 5.

Painting of the recall and precision for the recognition of the image for the linear kernel nucleus.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	96.38	96.89
I2	76.42	79.11
I3	93.38	95.16

Note: e.2 Classification with SVM using RBF (radial basic function) nucleus.

Table 6.

Painting of the recall and precision for the recognition of the image for the RBF kernel nucleus.

Image classified	Precision RBF kernel (%)	Recall for RBF kernel (%)
I1	97.29	97.96
I2	77.89	80.65
I3	94.45	96.13

Note: e.3 Classification with SVM using the polynomial kernel nucleus.

Table 7.

Painting of the recall and precision for the recognition of the image for the polynomial kernel nucleus.

Image classified	Precision polynomial kernel	Recall for l polynomial kernel
I1	95.56	95.97
I2	75.72	78.56
I3	92.79	94.45

Note: e.4 Classification with SVM using the sigmoid nucleus.

Table 8.

Painting of the recall and precision for the detection of linear behavior for the core.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	94.89	94.99
I2	74.91	77.83
I3	91.94	93.66

Figure 12:

Classification with SVM using the linear kernel nucleus

Figure 12 and Table 5.

Classification with SVM using RBF (radial basic function) nucleus

Figure 13 and Table 6.

Figure 13:

Classification with SVM using the polynomial kernel nucleus

Figure 14 and Table 7.

Figure 14:

Figure 15:

Classification with SVM using the sigmoid nucleus

Figure 15 and Table 8.

According to Figures 12-15 and Tables 5-8, we can say that the use of the RBF core for the ‎phase of learning gives the best results, it means the highest rate of the Image classification. This is due to the good choice of our database, and the methods used for this ‎classification, which are the Scal IFT for image parameterization, and the SVMs for learning and ‎classification. For the database, we register in the intelligent building the presence of a person, the sleeping person, and the absence of a person. These experiments are repeated 6 times, so that each experiment has a different lighting from the others, in order to take into account the variation in lighting. We also used different persons (age, sex), to play the scenario that represents a person present, a sleeping person, and an absent person.

The classification rate varies between 97 and 99.3% and with an accuracy that varies between 77.89 and 97.29% and a recall that varies between 80.65 and 97.96%. Images that have lower accuracy and recall have bad lighting, especially all images that express a sleeping person. The images express the presence of a person with the best precision and recall, because in this case we have a very good lighting.

The comfort command system in an intelligent building that we have created, is composed of two parallel systems, which work at the same time.

The first system consists in classifying the word, precisely the classification of 20 words; the second system consists of classifying the images, specifically the classification of three images (Fig. 16). Our system is robust and optimal, because we obtained good results for the precision, the recall, and the rate of the classification, and also, because we used only 20 words for the first system and three images for the second system, which represents an optimal number for the command of comfort. The good results obtained for the recognition rate, the precision, and the recall for both systems, are the condition to be taken into account for reliability and accuracy in this methodology.

Figure 16:

Comparative analysis

For the speech classification system

In order to validate our obtained results, a comparative study is essential, in order to better see the contribution of the MFCCs of the SVMs for the classification of speech. We repeated the same experiments with the same conditions, but we changed the method of the parameterization of words, such as we used the Linear PC (linear prediction coding) method. In this method, each word will be divided into 21 segments. Each segment will undergo a 23-ms HAMMING windowing (256 samples) with a 40% overlap. From each part 12 Linear PC coefficients are extracted. Order 12 is sufficient to characterize a part. After concatenation of the different blocks of results, each word will be represented by 252 coefficients (21 segments coded with 13 coefficients). W_{i Mel FCC} represents the word w by the Mel FCC method, and W_{i j} the coefficients.

W _{1 1} W _{1 2} …………………………………………W _{1 21}

W_{2 1} W _{2 2} ………………………………..……….. W _{2 21}

W_{3 1} W _{3 2} ………………………………...………..W _{3 21}

W _{4 1} W _{4 2} …………………………………..……..W _{4 21}

W _{5 1} W _{5 2} ……………………………………..…..W _{5 21}

W _{6 1} W _{6 2} ……………………………………..…..W _{6 21}

W_{i Linearl PC} = W _{7 1} W _{7 2} ………………...……….. W _{7 21}

W _{8 1} W _{8 2} ……………………………….………. W _{8 21}

W _{9 1} W _{9 2} …………………………………..……..W _{9 21}

W _{10 1} W _{10 2} ………………………………………W _{10 21}

W _{11 1} W _{11 2} ………………………………………W _{11 21}

W _{12 1} W _{12 2} ………………………………………W _{12 21}

We present the results obtained by the Linear pc method and SVMs for the classification of speech in Figures 17-24 and Tables 9-12.

Table 9.

Painting of the recall and precision for the recognition of the words for the linear kernel nucleus.

Word classified	Precision linear kernel	Recall for linear kernel
W1	90.36	91.75
W2	88.63	89.93
W3	93.75	95.13
W4	83.47	85.59
W5	54.23	62.19
W6	50.17	69.21
W7	81.01	85.01
W8	57.21	56.17
W9	77.69	83.65
W10	75.11	80.09
W11	79.63	87.67
W12	68.82	77.18
W13	75.56	81.45
W14	72.86	77.63
W15	77.32	85.17
W16	65.83	60.14
W17	88.93	91.04
W18	87.01	90.01
W19	91.72	94.72
W20	81.31	85.03

Table 10.

Painting of the recall and precision for the recognition of the words for the RBF (radial basic function) nucleus.

Word classified	Precision for RBF kernel (%)	Recall for RBF kernel (%)
W1	91.59	93.31
W2	90.17	91.39
W3	96.21	96.59
W4	85.76	88.13
W5	56.11	63.69
W6	52.56	70.51
W7	82.45	86.59
W8	59.32	57.72
W9	82.51	86.17
W10	80.23	82.89
W11	84.17	89.52
W12	70.46	79.31
W13	78.14	84.21
W14	74.86	80.14
W15	79.53	87.59
W16	67.54	62.45
W17	89.69	92.45
W18	87.79	90.13
W19	93.72	98.14
W20	84.09	87.75

Table 11.

Painting of the recall and precision for the recognition of the words for the polynomial kernel nucleus.

Word classified	Precision polynomial kernel (%)	Recall for polynomial kernel (%)
W1	89.51	90.84
W2	87.32	89.01
W3	92.50	94.15
W4	83.11	84.71
W5	53.11	61.21
W6	49.06	68.25
W7	79.62	84.14
W8	56.01	55.19
W9	76.48	82.91
W10	74.02	79.18
W11	78.70	87.02
W12	67.92	76.21
W13	74.52	80.56
W14	71.93	76.63
W15	76.49	84.26
W16	65.03	59.19
W17	88.01	90.12
W18	86.12	89.03
W19	90.56	93.89
W20	80.48	84.08

Table 12.

Painting of the recall and precision for the recognition of the words for the sigmoid kernel nucleus.

Word classified	Precision for sigmoid kernel (%)	Recall for sigmoid kernel (%)
W1	88.70	90.21
W2	86.51	88.27
W3	91.75	93.39
W4	81.60	83.90
W5	52.36	60.39
W6	48.32	67.40
W7	78.83	83.31
W8	55.32	54.30
W9	75.72	82.01
W10	73.34	78.35
W11	77.90	86.28
W12	67.08	75.39
W13	73.72	79.69
W14	71.07	75.86
W15	75.75	83.39
W16	64.32	58.40
W17	87.28	89.27
W18	85.29	88.25
W19	89.72	93.07
W20	79.75	83.24

Figure 17:

Classification with SVM using the linear kernel nucleus

Figures 17 and 18 and Table 9.

Figure 18:

Classification with SVM using RBF (radial basic function) nucleus

Figures 19 and 20 and Table 10.

Figure 19:

Figure 20:

Classification with SVM using the polynomial kernel nucleus

Figures 21 and 22 and Table 11.

Figure 21:

Figure 22:

Classification with SVM using the sigmoid nucleus

Figures 23 and 24 and Table 12.

Figure 23:

Figure 24:

From the results obtained in Figures 17-24 and Tables 9-12, we can say that our results obtained are better than the results obtained with the Linear PC method and the SVMs.

For the image classification system

A comparative study is essential to validate our obtained results for the image classification that uses Scal IFT and SVM methods. We conducted again the same experiments with the same conditions, but we changed the method of the parameterization of images, and we used two different methods, which are Local BP (Local Binary Pattern) histograms and Red GB (red green blue) color histograms.

Local BP histograms

We then apply to each image the Local BP (Local Binary Pattern) method for the configuration of these images in order to represent each image by a vector of the same size containing 64 values. The implementation takes place in two phases: learning and classification.

I_{i Local PC} represents the image I_i by the Local BP method, and I_{i j} the coefficients.

I _{1 1} I _{1 2} ………………………………….…………… I _{1 8}

I _{2 1} I _{2 2} …………………………………….………….I _{2 8}

I _{3 1} I _{3 2} ………………………………………….…… I _{3 8}

I_{i Local} BP=I _{4 1} I _{4 2} ……………………………………I _{4 8}

I _{5 1} I _{5 2} ………………………………………………. I _{5 8}

I _{6 1} I _{6 2} ………………………………………………..I _{6 8}

I _{7 1} I _{7 2} ………………………………………………..I _{7 8}

I _{8 1} I _{8 2} ………………………………………………..I _{8 8}

We present the results obtained by the Local BP method and SVMs for the classification of image in Figures 25-28 and Tables 13-16.

Table 13.

Painting of the recall and precision for the recognition of the image for the linear kernel nucleus.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	95.90	96.41
I2	75.91	78.60
I3	92.80	94.62

Note: e.2 Classification with SVM using RBF (radial basic function) nucleus.

Figure 25:

Classification with SVM using the linear kernel nucleus

Figure 25 and Table 13.

Classification with SVM using RBF (radial basic function) nucleus

Figure 26 and Table 14.

Table 14.

Painting of the recall and precision for the recognition of the image for the RBF kernel nucleus.

Image classified	Precision RBF kernel (%)	Recall for RBF kernel (%)
I1	96.70	97.41
I2	77.32	80.08
I3	93.85	95.58

Note: e.3 Classification with SVM using the polynomial kernel nucleus.

Figure 26:

Classification with SVM using the polynomial kernel nucleus

Figure 27 and Table 15.

Table 15.

Painting of the recall and precision for the recognition of the image for the polynomial kernel nucleus.

Image classified	Precision polynomial kernel (%)	Recall for l polynomial kernel (%)
I1	95.01	95.38
I2	75.14	78.02
I3	92.19	93.88

Note: e.4 Classification with SVM using the sigmoid nucleus.

Table 16.

Painting of the recall and precision for the detection of linear behavior for the core.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	94.37	94.41
I2	74.39	77.28
I3	91.40	93.04

Figure 27:

Figure 28:

Classification with SVM using the sigmoid nucleus

Figure 28 and Table 16.

From the results obtained in Figures 25-28 and Tables 13-16, we can say that our results obtained are better than the results obtained with the Local BP method and the SVMs.

Red GB histograms

We then apply to each image the Red GB (Red green blue) method for the configuration of these images in order to represent each image by a vector of the same size containing 256 values. The implementation takes place in two phases: learning and classification.

I_{i Red GB} represents the image I_i by the Red GB method, and I_{i j} the coefficients.

I _{1 1} I _{1 2} …………………………………...………… I _{1 16}

I _{2 1} I _{2 2} ………………………………………...…….I _{2 16}

I _{3 1} I _{3 2} …………………………………...………… I _{3 16}

I _{4 1} I _{4 2} ………………………………………………I _{4 16}

I _{5 1} I _{5 2} ………………………………..……………. I _{5 16}

I _{6 1} I _{6 2} ………………………………..……………..I _{6 16}

I _{7 1} I _{7 2} ………………………………………..……..I _{7 16}

I_{i Red GB} = I _{8 1} I _{8 2} ………………………...…………..I _{8 16}

I _{9 1} I _{9 2} …………………………………………...… I _{9 16}

I _{10 1} I _{10 2} …………………………………………….I _{10 16}

I _{11 1} I _{11 2} …………………………………………… I _{11 16}

I _{12 1} I _{12 2} ……………………………………………I _{12 16}

I _{13 1} I _{13 2} ……………………………..……………. I _{13 16}

I _{14 1} I _{14 2} …………………………………...………..I _{14 16}

I _{15 1} I _{15 2} ………………………………...…………..I _{15 16}

I _{16 1} I _{16 2} ……………………………………...……..I _{16 16}

We present the results obtained by the Red GB method and SVMs for the classification of image in Figures 29-32 and Tables 17-20.

Table 17.

Painting of the recall and precision for the recognition of the image for the linear kernel nucleus.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	95.42	95.96
I2	75.49	78.29
I3	92.47	94.28

Note: e.2 Classification with SVM using RBF (radial basic function) nucleus.

Table 18.

Painting of the recall and precision for the recognition of the image for the RBF kernel nucleus.

Image classified	Precision RBF kernel	Recall for RBF kernel
I1	96.40	96.99
I2	76.98	79.89
I3	93.59	95.24

Note: e.3 Classification with SVM using the polynomial kernel nucleus.

Figure 29:

Classification with SVM using the linear kernel nucleus

Figure 29 and Table 17.

Classification with SVM using RBF (radial basic function) nucleus

Figure 30 and Table 18.

Figure 30:

Classification with SVM using the polynomial kernel nucleus

Figure 31 and Table 19.Classification with SVM using the sigmoid nucleus

Table 19.

Painting of the recall and precision for the recognition of the image for the polynomial kernel nucleus.

Image classified	Precision polynomial kernel (%)	Recall for l polynomial kernel (%)
I1	94.78	95.01
I2	74.84	77.71
I3	91.90	93.69

Note: e.4 Classification with SVM using the sigmoid nucleus.

Table 20.

Painting of the recall and precision for the recognition of the image for the sigmoid kernel nucleus.

Image classified	Precision linear kernel (%)	Recall for linear kernel (%)
I1	93.97	94.02
I2	73.05	77.05
I3	91.11	92.89

Figure 31:

Figure 32:

Figure 32 and Table 20.

From the results obtained in Figures 29-32 and Tables 17-20, we can say that our results obtained are better than the results obtained with the Red GB method and the SVMs.

Comfort control is seen as a solution to optimize energy in the intelligent building. There are several methods to control this comfort, among these methods, we have the classification of speech, which gives the person who occupies the intelligent building the possibility of obtaining the desired comfort using his voice, and in this way the comfort will be optimized. But the disadvantages of speech classification are (1) if the person is sick and cannot speak, then in this case the speech classification system will not be able to command comfort. Energy will not be optimized. (2) If the person sleeps leaving the lights on, the television, or music on, then in this case, energy will not be optimized, because these equipments will not be off as long as that person is asleep. In order to overcome these drawbacks, we proposed to add another system of image classification in parallel to the system of speech classification.

This image classification system will be used to classify three images that express the presence of the person, a sleeping person, and the absence of the person. When the person is sick and cannot speak, then in this case, the image classification system will detect the presence of the person, and the intelligent building comfort management system will give an optimal comfort to the person. This person must be satisfied with this optimal comfort, because the speech classification system will not recognize his voice.

In the event that the image classification system does not detect the presence, the building comfort management system turns off all lights and other equipment in that room, and lowers the heating or air conditioning.

In the event that the image classification system detects that the person is asleep, then the intelligent building comfort management system turns off all lights and other equipment in that room.

It can be said that the addition of this image classification system has solved the drawbacks of the automatic speech recognition system. But the image classification system has some drawbacks, such as bad lighting, darkness, and shadow problem. The automatic voice recognition system will solve these drawbacks, such as allowing the occupant of the smart building to stay in the dark, or have poor lighting while still being comfortable.

In the experiments carried out, we have used for the two systems, 4 cores (linear, RBF, polynomial, and sigmoid) in order to choose the best model that gives the best results for the classification. The speech and the image classification system have given satisfactory results especially for the RBF kernel as for the speech classification system the classification rate varies between 96.4 and 99.41% and with a precision that varies between 53.11 and 96.96% and a recall that varies between 58.45 and 98.85%. For the image classification system, the classification rate varies between 97 and 99.3% and with an accuracy that varies between 77.89 and 97.29% and a recall that varies between 80.65 and 97.96%.

In conclusion, we can say that our goal has been achieved because we have realized a robust and efficient system for the control of comfort in the intelligent building.

From the comparative study that we carried out, we can say that the experimental results obtained are satisfactory for the methods used, which allows us to say that Mel FCC and SVM are recommended for the classification of speech, and Scal IFT and SVM are recommended for image classification, and therefore these methods are recommended for comfort command in an intelligent building.