Evaluation Model of College English Multimedia Teaching Effect Based on Deep Convolutional Neural Networks

With the acceleration of global integration, the demand for English instruction is increasingly rising. On the other hand, Chinese English learners struggle to learn spoken English due to the limited English learning environment and teaching conditions in China. *e advancement of artificial intelligence technology and the advancement of language teaching and learning techniques have ushered in a new era of language learning and teaching. Deep learning technology makes it possible to solve this problem. Speech recognition and assessment technology are at the heart of language learning, and speech recognition technology is the foundation. Because of the complex changes in speech pronunciation, a large amount of speech signal data, the high dimension of speech characteristic parameters, and a large amount of speech recognition and evaluation computation, the large volume of speech signal processing requires higher requirements of hardware and software resources and algorithms. However, traditional speech recognition algorithms, such as dynamic time-warped algorithms, hidden Markov models, and artificial neural networks, have their advantages and disadvantages. *ey have encountered unprecedented bottlenecks, so it is difficult to improve their accuracy and speed. To solve these problems, this paper focuses on evaluating the multimedia teaching effect of college English. A multilevel residual convolutional neural network algorithm for oral English pronunciation recognition is proposed based on a deep convolutional neural network. *e experiments show that our algorithm can assist learners in identifying inconsistencies between their pronunciation and standard pronunciation and correcting pronunciation errors, resulting in improved oral English learning performance.


Introduction
e demand for English learning [1] in China is increasingly rising due to global integration and China's increasing degree of internationalization. e tremendous Chinese pronunciation characteristics and the difference with English pronunciation with time and location constraints cause the lack of a domestic English learning environment. Also, good English teachers and standard classroom teaching [2] cannot fulfill the English learning needs for various reasons.
ese reasons caused comprehensive English teaching and learning to be a big problem for the people. English as a second language has become a study hotspot in the field of education. Artificial intelligence (AI)-enabled [3,4] learning has solved this problem with the advancement of computer science [5] and technology [6] and improvements in language teaching [7,8] and learning methods. is technology will disrupt the current language teaching and learning environment, allowing learners to learn independently and in any place. It will provide learners with reliable, objective, and timely pronunciation feedback and direct them; learners will also benefit from repeated listening to determine the difference between its pronunciation and standard pronunciation, correct their pronunciation errors, and improve their language learning performance [9][10][11][12].
e key to evaluating college English multimedia teaching [13,14] is the recognition of spoken English. It refers to the technology of converting speech signals [15] into corresponding objects or texts by automatic recognition using AI and machine learning [16][17][18][19]. Speech recognition and assessment technologies have advanced rapidly in recent years, thanks to advances in deep learning, big data, and cloud computing technologies. As a result, it can model human neurons to interpret data through multilayer depth transmission, which has been verified in speech recognition, and deep neural networks (DNNs) display outstanding advantages in solving complex problems. e rapid growth of graphing calculators and cloud computing technology has reduced the computational complexity of DNN [20,21]. As a result, research on multimedia teaching and its assessment using technologies focused on deep learning will significantly increase English class teaching abilities.
On the one hand, most domestic English learners use portable devices to aid in oral English learning, such as language repeaters [22], MP3 players, and cell phones [23]. However, these tools do not conduct voice recognition and are restricted to inquiry and follow-up reading. e role, however, is unable to provide learners with a fair and objective pronunciation evaluation and feedback. On the other hand, in evaluating oral English, the current oral English test is still based on manual scoring with a strong subjective will, different standards, and slow speed. e problem is due to differences in the knowledge structure and experience of scoring experts and even differences in the decisions of the same expert [11,[24][25][26][27]. ere are deviations in the evaluation of the same pronunciation. Subjective factors such as this lead to poor repeatability and stability of manual evaluation. Moreover, the use of manual scoring will consume a lot of manpower and material resources. Based on the above observations, this paper focuses on evaluating college English multimedia teaching effects as the primary research content. erefore, in this paper, a multilevel residual convolutional neural network is proposed. e proposed scheme is used to recognize spoken English pronunciation based on the deep convolutional neural network [28][29][30][31][32]. e proposed algorithm has been tested, can help learners distinguish between their own and standard pronunciation, correct pronunciation errors, and improve the efficiency of oral English learning. e paper's key contributions are as follows: (1) Based on a deep convolutional neural network, a multimedia-based English teaching [33,34] impact evaluation model is proposed. It helps learners distinguish between their own and standard pronunciation, correct pronunciation errors, and increase the quality of oral English learning. (2) is paper proposes a novel multilevel residual convolutional neural network, making up for the missing features to improve the recognition rate. e rest of the paper is organized as follows. In Section 2, related work is studied. In Section 3, the methodology is given, whereas results and discussion are explained in Section 4. Finally, Section 5 concludes the paper.

Related Work
Different aspects of related work have been studied in this section, such as language recognition, speech processing, and feature extraction mechanisms.

Spoken Language Recognition Process.
e general process of speech recognition [35,36] is shown in Figure 1. First, the computer's sound card is used to digitize the voice analog signal and collect the voice signal. According to the Nyquist sampling theorem, in the process of analog/digital signal conversion, the sampling frequency f s max is greater than 2 times the highest F max in the signal, as shown in the following equation: (1) en, the sampled digital signal can express the adequate information in the original speech signal more completely. Since the frequency of everyday speech is generally between 40 and 4000 Hz, the sampling frequency is set as 8 kHz in this paper.
en, the obtained speech signal is preprocessed, which includes preweighting, framing, winding, and endpoint detection. en, the characteristic parameters of the preprocessed speech signal are extracted. Finally, the speech feature parameters can be selected for model training or pattern matching.

Spoken Speech Signal Preprocessing.
Under the influence of oral and nasal radiation and glottis excitation, the highfrequency end of the average power spectrum of the speech signal is attenuated by 6 dB/OCT at about 800 Hz. erefore, before the speech signal analysis, a 6 dB/OCT high-frequency lifting preweighted digital filter is generally adopted to enhance the high-frequency part of the speech signal so that the spectrum of the speech signal becomes flat. e spectrum of the whole frequency band from low frequency to high frequency can be obtained with the same signal-tonoise ratio. e calculation equation of the filter response function is as follows: where z is the pre-emphasis coefficient, which is taken as 0.9375 in this paper. In this way, the result y(n) after preemphasis processing can be expressed by the input speech signal x(n) as follows:

Feature Extraction.
Voice signal function parameter extraction aims to remove redundant data that are not important to voice processing and analyze and process the voice signal. e original speech signal has a large amount of data. It has too much information that interferes with the semantics due to the difference of the speakers, the loudness, and the length of the sound. Hence, it is not suitable for direct use in speech processing. e quality of feature parameters directly impacts speech processing efficiency, and a suitable feature extraction method will yield better results. As a result, function parameters from the original voice signal must be extracted. e ideal voice function describes only semantic information, and the total amount of voice data is also tiny.

Methodology
In this section, different stages of the methodology are discussed. e details about the convolutional neural network and the proposed multilevel residual convolutional neural network model are elaborated.

Convolution Neural Network.
A convolution neural network (CNN) is put forward based on sparse interaction, parameters, share, and the critical thought of three identical mappings. It works through the set the convolution kernel size, and step length can be implemented in a small area in the image. In this way, the local characteristics of the image are extracted. Setting the pooling layer to dimension reduction of convolution image effectively reduces the number of network weights and improves the training efficiency. CNN will extract features at low, mid, and high levels in the image. Each convolution kernel in the convolution layer is equivalent to a feature extractor. e more the layers in a network there are, the more the features that can be deleted at different levels. e structure of CNN is shown in Figure 2. A convolutional layer, an activation layer, a pooling layer, a completely connected layer, and an output layer are usually included in a CNN. CNN's central component is the convolution layer. Feature extraction at different levels can be achieved by setting the convolution kernel. e calculation equations between CNN convolution kernels are as follows: where y l represents the output of the l-th convolutional layer, w l represents the tensor of the convolution kernel in the l-th convolutional layer, y l−1 represents the convolution output of the l − 1-th convolutional layer, and b l represents the bias of the l-th convolutional layer. e down-sampling layer is also known as the pooling layer. e pooling layer will reduce the input function map's dimensionality and remove the key features, minimizing overfitting to some degree. e average pooling method adds up all the values in the window to average. It uses the average as the final sampling value. e calculation process is shown in Figure 3(a). e maximum pooling method takes the maximum value in each window as the final. e sampling value and the calculation process are shown in Figure 3(b). It is worth noting that the maximum pooling method is used. e completely linked layer is the essential structure for classifying the high-level feature data obtained by the convolutional and pooling layers. e input feature map after convolution and pooling operations is compressed into a one-dimensional matrix, and then the one-dimensional matrix is input to the fully connected layer for training to realize the learning and memory of the target information. A large number of end-to-end neurons connect the fully connected layer. Each neuron is equivalent to a memory unit. Pattern recognition and classification can be realized through reasonable parameters such as weights and biases. e training process optimizes each neuron's parameters and bias values by forwarding propagation and back propagation algorithms and determining the optimal parameters.

Multilevel Residual Convolutional Neural Network Model.
CNN can handle grid data such as images well. Each convolution kernel can extract information such as image texture and edge features from different levels and improve the recognition efficiency by increasing the number of convolution layers. However, with the deepening of the convolutional layer, the extracted feature information becomes less and less semantic. Problems such as loss of original feature information are prone to occur, resulting in slower training convergence speed and difficulty in improving the recognition rate. Based on the above analysis, this article improves the typical CNN structure and designs a multilevel residual convolutional neural network, as shown in Figure 4. e proposed multilevel residual convolutional neural network contains multiple convolutional pooling layers and Mobile Information Systems a multilevel residual structure. e multilevel residual structure can transmit original information across multiple convolutional layers to compensate for missing features. It represents the number of convolutional layers that the residual structure spans. e multilevel residual convolutional neural network designed in this paper further reduces the amount of calculation. It improves the recognition efficiency based on solving the shortcomings of the CNN structure. e proposed model is improved using the residual structure by connecting the original information of the first n convolutional layers with the current layer. e structure of the multilevel residual retains the original information to the maximum extent. It adjusts the dimension of the original input features by adding control coefficients. Hence, it can effectively improve the recognition rate and accelerate the convergence rate. e principle is shown in Figure 5.
Assuming that the input when the residual structure is derived is x i and the output after the residual is introduced is x i+n , the output of the multilevel residual structure is as follows: where α and β are control coefficients, which are used to limit the dimension of input features. Assuming that the loss function is C, the weight update formula with backpropagation is as follows: When the convolutional neural network is differentiated layer by layer, the weight w will gradually decrease or even approach zero. It will lead to the gradient update of backpropagation approaching zero, namely, the phenomenon of feature loss. e original feature information of the first n convolutional layers can be introduced by adding a multilevel residual structure to supplement the features to the greatest extent. Meanwhile, the control parameters A and B can reduce the feature dimension, accelerate the training speed, and improve the training effect.

Experiments and Results
In this section, a detailed discussion on the experimental setup and result is performed. e following subsections, experimental setup, evaluation methods, datasets, experimental results, and model performance testing and analysis are performed.

Experimental Setup.
is experiment uses the Tensor-Flow 2.0 toolkit and Matlab2018b to complete the construction of the network model, data preprocessing, and the realization of the training algorithm. e server platform configuration used for training is shown in Table 1.

Evaluation Methods.
e main focus of the intonation test is to see if the material information in the pronunciation sentence is complete and correct. e MFCC coefficients based on the human ear hearing model are used as intonation assessment parameters in this paper, and the speech recognition model is developed using a deep belief network for speech recognition to assess if the content is complete and accurate. To judge the pronunciation, the correlation coefficient of the standard sentence and the MFCC function of the input sentence are both determined. Both intonation assessment and input are integrated on English pronunciation accuracy if it is consistent and fluent. e term "speech speed" generally refers to the "pronunciation speed," which measures how quickly a speaker pronounces words. It can be expressed as the number of syllables N spoken in a unit of time T. It can be approximately calculated as the total speech length, including pauses. Since different speakers talk at different speeds, different people pronounce the same sentence differently depending on the length of the sentence. Furthermore, the speaker's emotional state affects speech tempo. For example, the pace of speech is generally slightly faster in anger and happiness than in a calm state. In contrast, the speed of speech is generally slower in the state of sadness. e length ratio A between the test sentence and the regular sentence is calculated in this paper using the speech rate evaluation based on the speech duration. e calculation equation is as follows: where Len std is the duration of the standard sentence and Len test is the duration of the test sentence.

4.3.
Data. e subjects of this paper are college students in our school, a total of 57 people, including 37 boys and 20 girls. Subjects were recorded by CoolEdit, a recording software, with a sampling rate of 16 kHz and 16-bit coding. e recording contains 10 sentences, all of which are commonly spoken English sentences.

Experimental Results.
We compared the DHMM, CDHMM, TDA-GTS, and KASWT methods in the same experimental setting to somewhat check the superiority of the algorithm in this paper. e comparison results of their recognition rates are shown in Table 2.
e recognition rate of the proposed model is 97.17 percent, which is higher than the above models, as shown in Table 2 and Figure 6. As a result, the algorithm presented in this paper is both rational and accurate. It can be used to assess the impact of college English multimedia instruction. Figures 7 and 8 show that the proposed model will reduce the loss function value below 0.2. At the same time, DHMM can only drop to about 0.5. e loss function values of the KASWT method drops to about 0.2 and stops converging. It demonstrates that our model performs better in terms of convergence. In general, this algorithm outperforms the compared methods in terms of convergence efficiency and speed in the experiment.

Conclusion
Large-scale voice signal processing necessitates more challenging software and hardware resources and algorithms due to the complex changes in voice pronunciation, a large amount of data in voice signals, the high dimensionality of voice function parameters, and a large number of calculations for voice recognition and evaluation. Traditional speech recognition algorithms such as dynamic time warping, hidden Markov models, and artificial neural networks each have their own set of benefits and drawbacks. ey have hit unheard-of bottlenecks, and it is impossible to boost their accuracy and pace anymore. In response to these issues, the emphasis of this article is on evaluating the effects of college English multimedia teaching. A multilevel residual convolutional neural network is proposed to assess spoken English pronunciation using. e proposed algorithm has been tested, which helps learners distinguish between their own and standard pronunciation, correct pronunciation errors, and increase the quality of oral English learning.

Data Availability
e data used to support the findings of this study are included within the article.  Mobile Information Systems 7