A Machine Learning Assessment System for Spoken English Based on Linear Predictive Coding

In the teaching of English, there is an increasing focus on practical communication skills. As a result, the speaking test component has received more and more attention from education experts. With the rapid development of modern computer technology and network technology, the use of computers to assess the quality of spoken English has become a hot topic of research in related elds at present. A machine learning assessment system based on linear predictive coding is proposed in order to achieve automatic scoring of spoken English tests. First, the principle of linear predictive coding and decoding is analyzed, and the traditional linear predictive coding and decoding algorithm is improved by using hybrid excitation instead of the traditional binary excitation. Second, the overall structure of the machine learning assessment system is designed, which mainly includes division into four modules: acoustic model acquisition module, speech recognition module, standard pronunciation transcription module, and decision module. en, the speech recognition module is implemented by an improved linear predictive speech coding method to acquire the feature parameters of the speech signal and generate the speech feature vector. Finally, the convolutional neural network algorithm is used to train the speech features so as to implement the acoustic model acquisition module. e experimental results show that the improved linear predictive speech coding method yields more natural and higher intelligibility speech signals. e designed machine learning evaluation system is able to accurately detect information about the quality of the learner’s pronunciation.


Introduction
e focus of modern English language teaching is on the development of students' general application skills, including listening and reading skills. Among these, speaking training and speaking assessment have received increasing attention. ere are generally two types of assessment for speaking tests: an automated assessment and a manual assessment by experts. With the continuous development of random computer technology, automated assessment of speaking tests is beginning to be used in a variety of industries [1][2][3][4][5][6]. For example, speaking assessment systems can be used during telephone interviews to automatically score the English pro ciency of interviewees. In addition, online teaching application scenarios in the education industry can use speaking assessment systems to automate the scoring of students' speaking quality. Automated speaking assessment systems can give objective scores based on the test taker's performance in a timely manner and are not subjectively in uenced by personal factors [7,8].
As competition in business continues to intensify, there is an increasing demand for complex talents. Companies require these people to have not only solid professional knowledge, but also to be able to express themselves prociently in English, so speaking skills are quite important. Unlike traditional written English teaching, oral teaching focuses on standard pronunciation. Although the forms of teaching have diversi ed, spoken English teaching is still at an arti cial stage at this stage. In the traditional language teaching process, teachers provide comprehensive training such as listening, reading, and writing to students through a face-to-face approach, so as to achieve the purpose of developing students' language communication skills [9][10][11]. Among them, the learning and training of standard spoken language is the foundation and focus of English learning. Due to the constraints of teachers' resources, learning costs, and learning locations, the effect of traditional speaking learning and training is not satisfactory. Teachers need to spend a lot of time and effort conducting various subjective tests on students, resulting in ineffective work efficiency, especially in large-scale speaking test scenarios.
Currently, researchers are beginning to experiment with computer-assisted pronunciation training systems to address these problems [12][13][14]. e core issue of computeraided pronunciation training systems is pronunciation bias testing, i.e., pronunciation bias assessment. Pronunciation bias assessment is the assessment of the standard of the learner's pronunciation and the assignment of a corresponding score or grade, which is the core function of a computer-aided pronunciation training system. Pronunciation bias assessment is mostly a confidence-based method. e phoneme sequence is first standardized and sliced to obtain more accurate phoneme boundary information.
en, the confidence of the phonemes in each speech segment is calculated, and the pronunciation bias is measured by the confidence score. Common confidence calculation methods include log-likelihood, log-likelihood ratio, logposterior probability, and Goodness of Pronunciation (GOP) [15][16][17]. In addition, some methods combine confidence calculations with pronunciation features, which yield better joint score results. In order to assess pronunciation bias with high accuracy, more and more researchers are focusing on the detection of pronunciation bias at the phoneme level.
ere are two ideas for the study of automatic detection of pronunciation bias at the phoneme level [18]. One is an automatic method for the detection of pronunciation bias based on acoustic phonetics. Such methods are based on a statistical analysis of speech. e other is an automatic method of pronunciation bias detection based on automatic speech recognition technology.

Pronunciation Bias Detection Based on Acoustic Phonetics.
Pronunciation bias detection based on acoustic phonetics finds a specific combination of features by extracting structural, acoustic, and perceptual features of the speech to be tested. en, pronunciation bias detection is achieved by statistically examining. A similarity calculation or a classifier is usually chosen for the differentiation of pronunciation bias types.
Morlett Paredes et al. [19] proposed a hybrid method based on time-domain features and phoneme boundary information for pronunciation bias detection of basic English pronunciation units, with remarkable results. is hybrid method used a multilayer perceptron as a classifier. Nakamura et al. [20] extracted several resonance peaks from different frames after pre-processing the speech to be measured. en, a Gaussian Mixture Model (GMM) was used for classification and vowel articulation bias detection was achieved. Dashti and Razjmoo [21] defined a resonance peak that reduces ambient noise. is resonance peak is able to simulate the vocal tract shape properties. Articulatory bias detection is then performed by calculating the degree of structural distortion (Bhattacharyya distance) between the speech to be measured and the standard speech.

Pronunciation Bias Detection Based on Automatic Speech
Recognition Technology. Automatic speech recognition is essentially a classification matching problem, while pronunciation bias detection is a classification regression problem, so pronunciation bias detection can be solved using speech recognition technology. Pronunciation bias detection based on automatic speech recognition is simpler than pronunciation bias detection based on acoustic phonetics. is is because automatic speech recognition can use a language model to counteract the effects of imprecise acoustics and thus output a legitimate sequence of characters. erefore, this study chose to use automatic speech recognition to implement a spoken English assessment system. e key elements of automatic speech recognition technology include the extraction of speech feature parameters and the selection of acoustic models, both of which are also the focus of this study.
First of all, the extraction of speech feature parameters is a key step in the process of dynamic speech recognition, and the selection of parameters directly affects the overall performance of the system. After the speech signal has been preprocessed, it needs to be extracted and analyzed for the feature parameters. e most typical method of extraction is the use of vocoders. e vocoder was born in the 1920s at Bell Labs in the USA. Since then, the vocoder has seen a period of rapid development. A large number of researchers have been working on speech coding and speech synthesis, and have achieved considerable results. e basis of the vocoder is Linear Predictive Coding (LPC). In the early 1980s, the US Department of Defense published LPC-10. Liu et al. [22] used LPC to build a parametric pronunciation bias database and combined it with a Gaussian Hidden Markov Model to achieve classification detection of pronunciation bias. Hiroya and Mochida [23] used LPC to extract speech feature parameters and then used the linear discriminant analysis or decision trees to train classification models to achieve pronunciation bias detection.
Second, for pronunciation detection, the constraint provided by the language model is not helpful as it leads to missed detection of incorrect pronunciations. erefore, robust acoustic models are important to distinguish between those with standard pronunciation and those with abnormal pronunciation. In traditional speech recognition, the Gaussian Hidden Markov Model (GMM-HMM) has been the dominant acoustic model [24]. However, with the continuous development of deep learning techniques, deep learning models are gradually being used more often in speech recognition tasks. A convolutional neural network (CNN) is a multilayer perceptron that incorporates convolutional computation. CNN is one of the representative algorithms of deep learning [25] and is commonly used to analyze visual images. e CNN consists of an input layer, a convolutional layer, a ReLU activation layer, a pooling layer, and a fully connected layer. e CNN is also known as a "translation-invariant artificial neural network." When applied to automatic speech recognition applications, in terms of input, CNN-based automatic speech recognition techniques are broadly divided into two types: one is to use traditional acoustic feature parameters as input, such as Mel Frequency Cepstrum Coefficient (MFCC) [26], LPC [27], and Fbank [28]. e other is to use original timefrequency spectrum as input, that is, to treat the time-frequency diagram as an image. Er et al. [29] analyzed the research on deep learning techniques in speech recognition and the key problems to be solved. Nakashika et al. [30] used recurrent neural networks for speech recognition and the recognition accuracy was high.
From a pronunciation bias detection perspective, we want to retain as much of the original information as possible in the features received at the input. is is because the original information is the most realistic representation of the quality of the learner's spoken language. However, time-frequency maps can cause information loss in the frequency domain, which is detrimental to pronunciation bias detection. erefore, the automatic speech recognition technology in this paper uses acoustic feature parameters as input information. Due to the short-time smoothness of spoken English, the feature parameters of the acoustic model in pronunciation bias detection are updated less frequently, which effectively reduces the coding bit rate (below 2.4 kb/s or even below). e simple LPC vocoder is able to achieve a range of 0.8 to 2.4 kb/s in terms of coding efficiency, which just meets the coding bit rate requirements [31][32][33]. erefore, LPC is used for speech signal feature extraction, and the features are trained by convolutional neural network algorithm to complete speech recognition. e aim of this study is to adopt LPC to extract acoustic feature parameters and use CNN as an acoustic model for pronunciation bias detection to automate the detection of English pronunciation bias.
In order to achieve automatic scoring of spoken English tests, a machine learning assessment system based on linear predictive coding is proposed, which mainly consists of being divided into four modules: acoustic model acquisition module, speech recognition module, standard pronunciation transcription module, and decision module. e improved stimulated linear predictive speech coding method is used to obtain the feature parameters of the speech signal and generate the speech feature vector to implement the speech recognition module. Finally, the CNN model is used to train the speech features so as to implement the acoustic model acquisition module. e experimental results show that the improved LPC + CNN-based evaluation system can accurately detect pronunciation bias information.
e main innovations and contributions of this paper include.
(1) How accurately unvoice/voiced tones judgments are made is important for spoken English assessment systems. erefore, the traditional LPC algorithm is improved by using hybrid excitation instead of simple binary excitation. In the acoustic feature parameter extraction process, the sub-band sound intensity of the speech signal is extracted using a split-band hybrid excitation technique in addition to the extraction of the fundamental tone period required by the traditional LPC model. (2) An English spoken pronunciation evaluation system based on improved LPC and CNN is constructed. e improved LPC algorithm is used to obtain the feature parameters of the speech signal and generate the speech feature vector, thus realizing the speech recognition module. A CNN is used to train the speech features, thus realizing the acoustic model acquisition module. e rest of the paper is organized as follows: In Section 2, the representative spoken pronunciation assessment system was studied in detail, while Section 3 provides the improved LPC algorithm. In Section 4, the machine learning evaluation system based on ILPC + CNN was studied in detail, while Section 5 provides experimental results and analysis. Finally, the paper is concluded in Section 6.

Representative Spoken Pronunciation Assessment System
Since the 1990s, many technology companies and research institutes have conducted in-depth research in the field of pronunciation bias testing and have achieved remarkable results, and launched various application systems, as shown in Table 1. ese systems have been widely used in areas such as computer-aided pronunciation training, computer-aided language learning, and computer-based speaking proficiency testing. For example, the DISCO (Development and Integration of Speech technology into Courseware for language learning) project at the University of Nijmegen (Netherlands) [34]. e DISCO system automatically detects pronunciation deviations and grammatical errors in the speech to be tested and generates detailed feedback on the errors checked. e HUGO system, developed by Kyoto University in Japan for Japanese learners of English, uses a decision tree technique based on linguistics and a phonological database to check pronunciation bias.

Principle of LPC.
e most basic low-rate speech coding method is linear predictive coding. In speech signal analysis linear prediction not only enables predictive functions but also provides a very good estimation of the vocal channel model parameters. Linear prediction analysis can provide a set of speech signal model parameters that accurately represent the spectral amplitude of the speech signal. e basic idea of linear predictive analysis is to use the p sample point values of the previous set of data to predict the sample point values of the current or next set. LPC can simulate the human articulatory system very well and therefore has some advantages in the extraction of English speech feature parameters [35]. After waveform interception and noise filtering of the speech signal, multiple frames of speech signal in a certain time period can be obtained by frame sampling and combined with a linear time domain model to achieve feature parameter extraction.
Let s(n) represent the speech signal. According to the LPC principle, s(n) can be represented by the previous p sample points.
where a 1 , a 2 , a p denote linear prediction coefficients. Let s(n) be the predicted speech signal, then its representation is shown as follow: e prediction error is calculated as shown as follow: , then all coefficients can be solved and a stable speech feature signal can be obtained. e basic principle of the linear predictive vocoder is that the model parameters are encoded with the excitation parameters using linear predictive analysis in an all-pole vocal channel model, resulting in the transmission of high-quality speech at low bit rates (below 2.4 kb/s). e principle of the linear predictive vocoder is shown in Figure 1. At the receiver end of the linear predictive vocoder, the prediction coefficients obtained from the linear predictive analysis can be used to synthesize the transmitted speech directly [36]. Figure 2 shows the coding principle of the LPC-10 vocoder.
First, after a low-pass analog filter, the LPC-10 vocoder performs an A/D conversion at a sampling rate of 8 kHz to obtain the digitized information of the speech. e digitized speech is then processed simultaneously in two steps. (1) e excitation information is processed. After the speech has been framed, the characteristic parameters of each frame are extracted and encoded for transmission. After encoding, the fundamental tone period (Pitch) and the voiced/unvoice sign (V/UV) of each frame are obtained. e fundamental tone period is calculated using the average amplitude difference function (AMDF) method. (2) e extraction of the vocal channel parameters is processed.
Because most of the energy of the speech signal is concentrated in the low-frequency range and the power spectrum decays with frequency, the LPC needs to preprocess the speech signal first so that the power spectrum on the high frequencies can be increased, thus improving the accuracy of speech channel parameter extraction.
where H pw (z) denotes the transfer function of the preprocessing filter.

3.2.
Improvements to Incentive Sources. Conventional LPC algorithms use simple binary excitation sources (voiced/ unvoice) to excite the synthesizer. Due to the low robustness, the quality of the speech synthesized by the binary excitation source is poor in the presence of high speech noise. Real-life English speech often has both voiced/unvoice tones, especially in noisy speech segments. erefore, the result of the voiced/unvoice tones judgment can directly affect the quality of speech recognition. erefore, improvement of the excitation source is important for spoken English evaluation systems. In this paper, a hybrid excitation is used instead of the traditional binary excitation, thus proposing an improved LPC algorithm (ILPC). In terms of parameter extraction, in addition to extracting the fundamental tone period required for conventional LPC, the hybrid excitation technique is also used to extract the sub-band sound intensity in the speech signal. e steps for extracting the fundamental tone period in ILPC arithmetic are shown as follows: Step 1: after passing the speech signal x(n) through a low-pass filter at 900 Hz, the first 20 output values are removed to obtain x′(n).
Step 2: find the maximum amplitude value of the first 100 samples and the maximum amplitude value of the last 100 samples of x ′ (n) respectively. Select the smallest value as the threshold level L. Step 3: make center-decimation and three-level decimation ofx ′ (n) to obtain y(n) and y ′ (n) respectively.
Step 4: find the correlation R(k) between the signals y(n) and y ′ (n) ; where k ranges from 20 to 150 and R(0) is the shorttime energy.
Step 5: use the peak detector to find the maximum value of the correlation value R max . If R max is less than 0.25R(0), this frame is considered as voiced tones and the fundamental tone period is set to P � 0. Otherwise, this frame is considered unvoiced tones, and the fundamental tone period is set to P � k.
e process of extracting the sub-band sound intensity in the ILPC calculation is shown in Figure 3.
After passing through the bandpass filter, the speech signal is extracted to obtain the fundamental tone period. e result of passing a frame of speech signal through each of the five sub-band filters is shown in Figure 4. e sound intensity of the five sub-bands is calculated as follows: 0.2452 for the first sub-band; 0.4478 for the second sub-band; 0.1893 for the third sub-band; 0.3707 for the fourth subband; and 0.3874 for the fifth sub-band.
For each unvoiced tone frame or dithered turbulent frame, the sound intensity of the speech signal in each subband is calculated separately. In forming the excitation signal, the sound intensity will determine the weighting of the pulse and noise sources in each sub-band, resulting in an excitation signal for the entire frequency band.

General System
Architecture. e automatic detection of pronunciation bias is a simulation of the human subjective detection process. rough machine learning of the manual detection results, automatic detection can even outperform human experts. e machine learning evaluation system for spoken English designed in this paper is shown in Figure 5.
e system is divided into four modules: an acoustic model acquisition module, a speech recognition module, a standard pronunciation transcription module, and a decision module.

ILPC-Based Speech Recognition Module.
In this paper, the ILPC algorithm is used to implement a speech recognition module so that the learner's basic pronunciation units (phonemes), including legal and illegal pronunciation unit sequences, can be accurately identified. Automatic speech recognition aims to detect the content of the learner's pronounced text and to output legitimate character sequences by using acoustic models that can counteract the effects of undesirable acoustics. In the acoustic feature parameter extraction process, we use a split-band hybrid excitation technique to extract the sub-band sound intensity of the speech signal in addition to the parametric fundamental tone period, resulting in an accurate voiced/unvoiced tones judgment.

CNN-Based Acoustic Model Acquisition Module.
e main function of the acoustic model acquisition module is to train an acoustic model. e trained acoustic model will be used in the speech recognition module. In traditional speech recognition, the Gaussian Hidden Markov Model (GMM-HMM) has been the dominant acoustic model. However, with the continuous development of deep learning techniques, deep learning models are gradually being used more often in speech recognition tasks. A convolutional neural network (CNN) is a multilayer perceptron that incorporates convolutional computation [37]. CNN is one of the representative algorithms of deep learning and is commonly used to analyze visual images. erefore, in this paper, CNNs are used to implement the acoustic model acquisition module.   Mobile Information Systems e input to the CNN is the acoustic feature parameters obtained by the ILPC algorithm. the structure of the CNN is shown in Figure 6. Let the sample set of speech features be � (x 1 , x 2 , x N ). First, the m speech features are convolved in the layer l of the CNN [38][39][40].
where k lj and b l,j represent the weights and biases of the features j in thel layer respectively, and * represents the convolution operation.
en, the convolution operation is performed on the m features of the N samples. Let the size of the convolution kernel be h × w.
A new sample is obtained again after the convolution operation and a transformation operation is performed on it.
e restrictions are shown as follows: After obtaining the fully connected layer of the convolutional neural network, the classifier is selected to predict the sample class.
In traditional acoustic model training, the label corresponding to each frame of data needs to be known in order to train effectively. erefore, the speech signal needs to be forcibly aligned prior to training the model. Although there are some relatively mature open source alignment tools available, there are significant constraints on the performance of speech recognition techniques with forced alignment. In CNN-based acoustic models, we want to leave more tasks to the neural network to perform, such as learning how to align autonomously. erefore, predictive alignment techniques are used to solve this problem. e loss function for predictive alignment is defined as shown below [41][42][43].
where p(y | Y) denotes the probability when the input sequence is y and the output sequence is Y, and S denotes the training set. It can be seen that prediction alignment can directly output the predicted probability of a sequence without the need for external post-processing. With the help of predictive alignment, a large amount of manual resources can be saved, thus increasing the efficiency. In this paper, the acoustic model acquisition module is built by combining prediction alignment and CNN, as shown in Figure 7.

Experimental Results and Analysis
In order to verify the performance of ILPC + CNN in the quality assessment of spoken English, various experiments were conducted using separate speech samples with different accents. e experimental speech data were obtained from the open-source website VoxForge (https://www.voxforge. org/zh). e parameters of the experimental dataset are shown in Table 2, with a ratio of 3 : 1 between training and test samples. e parameters of the CNN model are set as shown in Table 3. e sampling rate of the audio data is 16000 Hz and the sample size was 16 bit. e number of texts (number of pronounced sentences) is 2268 and the total number of phonemes is 44359. e total number of speakers is 10 including 5 males and 5 females. First, the effect of different frame rates on ILPC performance was tested. Second, the effect of different convolutional kernel sizes on the recognition performance was tested. Finally, the designed system was compared with other spoken pronunciation evaluation systems.

Effect of Different Frame Rates on ILPC Performance.
In order to obtain the best frame rate setting, the speech recognition accuracy of the six datasets at different frame rates was verified, as shown in Table 4.
It can be seen that as the number of frames extracted increases, the recognition accuracy keeps improving. As the frame rate increases to 180 Hz, the ILPC algorithm shows a high recognition accuracy. As the frame rates increases to 200 Hz, datasets 1, 2, 3, and 6 show a decrease in recognition accuracy, while datasets 4 and 5 show a very small and almost negligible increase in recognition accuracy. When ILPC was used to extract features of speech signals with different sample types, too high a frequency would increase the computational effort of speech recognition, while too low a frequency would drop important features of the speech signal. erefore, the frame rate used in the subsequent ILPC algorithms was 180 Hz.

Effect of Convolutional Kernel Size on Speech Recognition.
To further verify the effect of convolutional kernel size on speech recognition performance, the English speech recognition accuracy under different convolutional kernel conditions was tested, as shown in Table 5.
It can be seen that the recognition accuracy of English speech decreases when the size of the convolutional kernel increases.
is may be because fewer speech features are involved in the operation when the convolutional kernel size is too large, resulting in a decrease in speech recognition accuracy. e comparison shows that the recognition accuracy of CNN is higher when the convolutional kernel size is 2 * 2 and 3 * 3. However, 2 * 2 is more time-consuming than 3 * 3 in CNN operations, so to improve real-time performance, the convolutional kernel size was 3 * 3 in subsequent experiments.

Performance Analysis of ILPC.
e excitation signal of the ILPC vocoder is compared with that of the LPC vocoder, as shown in Figure 8.
It can be seen that the excitation signal obtained by ILPC is a mixed excitation signal. Each frame of speech is no longer pure unvoice tone or voice tone but contains a distinct periodic pulse string and a little noise. As a result,      the speech signal obtained by ILPC is more natural and better defined. Conventional LPC uses a simple binary excitation signal to process the input sequence. Compared to the conventional LPC algorithm, ILPC based on hybrid excitation gives a waveform that more closely resembles that of the original speech signal. ILPC algorithm can get speech signals with high naturalness, and its waveform is almost consistent with the original waveform.
To verify the performance of the ILPC-based speech recognition module, 1000 samples were taken from each of the six datasets to form a speech hybrid dataset containing 6000 samples. e spoken pronunciation bias of this hybrid dataset was examined using LPC + CNN and ILPC + CNN respectively. e frame rate was 180 Hz and the convolutional kernel was 3 * 3. e detection results are shown in Table 6 and Figure 9.
It can be seen that after ILPC feature extraction, the detection accuracy of CNN is significantly improved. Due to the lower robustness, the speech quality of LPC is poor in the case of very noisy speech, which is due to the fact that real-life English speech usually has both voice tones and unvoice tones, especially in transition segments and very noisy speech segments. When using ILPC for feature extraction of the captured speech signal, each frame of speech is no longer pure voice tones and unvoice tone, thus retaining as much of the original information as possible. ILPC + CNN converges at about 140 iterations, whereas LPC + CNN takes about 180 iterations to stabilize. In  Mobile Information Systems addition, the standard deviation of ILPC + CNN is smaller compared to LPC + CNN.

Performance Comparison of Different Spoken Language
Assessment Systems. In contrast to traditional spoken pronunciation assessment methods, the training data for this experiment did not require manual annotation. Using the above speech mixture dataset containing 6000 samples, the designed system was compared with other spoken pronunciation assessment systems, the results are shown in Table 7.
A total of 44359 phonemes (initials, finals, and tones) were obtained from the speech mixture dataset. e manual detection results showed that 6033 phonemes were mispronounced in this speech data, 10407 mispronounced phonemes were detected by the SCILL system and 9894 mispronounced phonemes were detected by the TBALL system. e system designed in this paper (ILPC + CNN) detected 7856 mispronounced phonemes, which is the closest to the manual (labeled) detection result. e experimental results show that ILPC + CNN algorithm can indeed reduce the misjudgment rate of pronunciation deviation. is indicates that the feature parameters obtained by ILPC using hybrid excitation reflect well the characteristics of the original speech signal and therefore the decoded speech quality is better and the speech is clearer.
Finally, the experiment classified the 64 pronunciation errors into three types, namely initial errors, final errors, and tone errors. ese three types of pronunciation errors were counted and the results are shown in Figure 10.
As you can see, of the 3 types of pronunciation errors, intonation is the most likely to occur. erefore, learners of English need to focus on intonation. e next problem is rhyme errors. Compared to the other two types of pronunciation errors, vowel errors are easier to solve. e phenomenon that tone errors are much higher than the other two types of errors is in line with linguistic laws and therefore the experimental results are reliable.

Conclusion
In this paper, a machine learning evaluation system for spoken English based on ILPC + CNN algorithm is constructed so as to automate the detection of learners' pronunciation errors. e designed system consists of four main modules: acoustic model acquisition module, speech recognition module, standard pronunciation transcription module, and decision module. e speech recognition module uses the ILPC algorithm to obtain the feature parameters of the speech signal and generate the speech feature vector. e acoustic model acquisition module uses a CNN model to train the speech features and the input to the CNN is the acoustic feature parameters obtained by the ILPC algorithm. e experimental results show that the feature parameters obtained by ILPC using hybrid excitation reflect the characteristics of the original speech signal very well, and therefore the decoded speech quality is better and the speech is clearer. Compared with other spoken English evaluation systems, the ILPC + CNN-based machine learning evaluation system can reduce the misjudgment rate of pronunciation bias.

Data Availability
e experimental data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e author declares that there are no conflicts of interest to report regarding the present study.