Application of Discriminative Training Algorithm Based on Intelligent Computing in English Translation Evaluation

For most Chinese schools and teachers, the teaching evaluation of spoken English is much more difficult than that of written English. Foreign language ability is a part of the national comprehensive quality, and the improvement of foreign language proficiency will help a nation gain more right to speak in the world. At present, English is the universal language in the world. Mastering English will have the tools to communicate with the world. The most important purpose of language learning is communication, and listening and speaking ability is the most important language skill [1].

In the early days, under the influence of the test-oriented education model, English teaching in China pays too much attention to the cultivation of reading and writing ability and neglected the training of oral expression ability. In English classes, teachers mainly talk about grammar and vocabulary, and there are few opportunities for oral expression. It can be said that oral language is basically not the focus of teaching, resulting in the common phenomenon of ‘dumb English’ and ‘Chinese English’ among Chinese students, which seriously affects the future development of students.

The difficulties in the teaching of spoken English in China can be summarised into three points. First, there are many students in large classes, and it is difficult for every student to have the opportunity to speak in class; second, there is no platform for students to practice oral English and get instant evaluation and feedback after class [2]; the organisation of the examination is very difficult. It requires a special voice computer room, supporting software and hardware computer test system, manual grading system and test question resources. Teachers need to spend several hours on correcting the written interview papers, in order to fully listen to the students' answering audio and give a rating. This situation is slowly changing. In recent years, in order to meet the needs of society and promote the reform of quality education, education authorities across China have actively explored the oral English test in the middle and high school entrance examinations. There are many types of questions in the oral English test of the middle and high school entrance examination, and the number of candidates is large. The traditional oral English test is conducted by manual face-to-face test or computer recording and manual scoring. This test method is difficult to organise and expensive to implement, and the scoring results are easily affected by the subjectivity of the scorers, which is not conducive to large-scale development.

Therefore, before 2011, the college entrance examination in each province only had an additional test for oral English, which was not included in the college entrance examination score. In 2011, Guangdong Province took the lead in not setting up an additional oral English test and directly included English listening and speaking scores into the total score of the college entrance examination. Candidates who take the English subject test must take the test, and the test results are obtained through manual double scoring. In 2014, the computer intelligent voice evaluation system developed by iFLYTEK was officially applied to the English listening and speaking test of the Guangdong college entrance examination. All the scoring work was completed within 2 days, which greatly reduced the difficulty of the organisation of the test and improved the scoring efficiency. The large-scale oral English test has become a reality, entering the era of computer intelligence evaluation, and it is feasible to promote it nationwide. At present, Canton phonetic evaluation technology helps oral English teaching, and voice evaluation technology helps oral English teaching and evaluation to be applied. Wenzhou and other 32 provinces and cities have included the English listening and speaking test into the senior high school entrance examination, and nearly 15 provinces and cities have included the results in the total score of the senior high school entrance examination. The computer intelligent scoring system of iFLYTEK has been widely used in auxiliary scoring.

Generally speaking, the measurement of confidence in the objective assessment of English pronunciation is obtained by calculating the correlation with the subjective assessment score. A high correlation coefficient indicates that the computer objective evaluation result is closer to the expert subjective evaluation, and the objective evaluation score is more reliable and more credible to reflect the quality of speech pronunciation. The low correlation coefficient indicates that the objective evaluation results of the computer are far away from the subjective evaluation results, so the reliability of the objective scores obtained by the computer is judged to be low [3]. In order to obtain the objective evaluation score with high confidence, the traditional method is to obtain the pronunciation quality evaluation score goodness of pronunciation (GOP) of each speech frame through the forced matching algorithm, and the GOP score is weighted and mapped to the GOP of phonemes and words. The purpose is to obtain objective scores that are more consistent with subjective evaluations [4].

Speech signal is a typical time-varying signal [5]. When the observation time is reduced to a small enough range, a series of approximately stable signals can be obtained. Therefore, when analysing the speech signal, we always assume that the speech signal is stationary within a time frame, which is the short-term analysis assumption of the speech signal. Usually a frame is about 20 ms or so. A set of features can be obtained by windowing a frame of signals and then performing feature analysis, and then processing the next frame by shifting the analysis window by an offset. There are related Hamming window or rectangular window functions in MATLAB speech signal processing. Practice has proved that this short-term stationary assumption is effective for the extraction of speech signal feature vectors. Feature vector extraction is an important part of speech evaluation. Feature vector extraction solves the digital representation of time-domain speech signals and provides data for the regression part.

Feature vector extraction is an important link in speech evaluation [6]. Feature vector extraction addresses the digital representation of time-domain speech signals [7] and provides data for the regression part. The quality of feature extraction directly affects the performance of the regression machine.

There are various methods for extracting speech features, such as ZCR, LPC, LPCC and MFCC. The linear prediction coefficient (LPC) method has been successfully applied in speech recognition due to its robustness to environment and speaker changes. In this study, the LPC method is used to process the sound signal frame by frame, and the linear prediction cepstral coefficients (LPCCs) are used as the feature vector of audio information. The LPC uses a finite number of parameter models to linearly approximate the audio sequence x(n). These parameters are called LPCCs. The extraction process of the LPCC is as follows: first, pre-emphasise the sampling points contained in each audio frame, add a window function to the pre-emphasised audio intra-frame signal and then perform autocorrelation analysis on it, and apply this result to p. In order linear prediction calculation, the sequence of length p is obtained, which is the LPC-derived cepstral coefficient of the audio frame; the linear prediction cepstral feature is extracted from each short-term audio frame, which mainly reflects the short-term frame characteristics of the audio. In this study, the 10th-order LPCC is used as the feature vector of each frame of speech [8], and some other features can also be added according to the situation. The speech feature extraction process is shown in Figure 1.

Fig. 1

In speech recognition, the maximum likelihood criterion is often used for the training of acoustic models [9, 10, 11], but the maximum likelihood criterion only considers increasing the likelihood of the correct category and does not consider the information of misclassification, which may cause problems in improving the correct category. At the same time as the similarity, the likelihood of the wrong category becomes higher. The discriminative training is not aimed at maximising the likelihood of the training corpus but focusses on how to adjust the classification surface between different models to improve the recognition ability of the model. In recent years, the discriminative training has significantly improved the performance of the speech recognition system. The representative training criteria are the maximum mutual information, minimum classification error and minimum word error. In order to further improve the generalisation of the model to improve the performance and robustness, the method similar to the large decision margin in machine learning is applied to the field of speech recognition to increase the distance between the correct recognition result and the wrong recognition result. For example, the maximum margin estimation criterion, the soft margin estimation criterion and enhanced confusing information methods mainly include enhanced maximum mutual information/minimum phoneme error and discriminative combinations between different models.

The data selection method based on the minimum phoneme error criterion proposed by Liu et al. [12] is based on the entropy value of the posterior probability of the Gaussian distribution in the state and selects training samples in the unit of time frame. A method of selecting data such as training sentences and candidate arcs in the expected phoneme accuracy field [13]. However, the aforementioned data selection methods are not sufficiently related to the calculation of statistics required for the update process of the acoustic model. In order to improve the combination of data selection and statistics and the efficiency of discriminative training, this study adopts the dynamic weighting method, which combines the posterior probability and the expected phoneme accuracy rate field to select training samples and competing candidates. The discriminative training algorithm focusses on the contributing samples [14, 15, 16] to adjust the parameters of the acoustic model. This study firstly selects candidate paths and candidate arcs that contribute more to the model statistics in the posterior probability word graph, based on the error of the candidate paths. Then, according to the confusion information of the phoneme pair, the penalty weight is applied to calculate the phoneme accuracy. On this basis, the distribution of the expected phoneme accuracy of the candidate arcs is estimated, and the candidate arcs are weighted. Finally, the performance of the proposed method is discussed [17, 18, 19, 20].

Discriminative algorithms generally use word graphs to describe the competitive relationship between candidate words and calculate and adjust acoustic model statistics based on word graph information. The word graph contains many candidate paths, which effectively describe the correct recognition results and candidate words. The competitive relationship between them provides sufficient confusion information for discriminative training. Given a training sentence z, the objective function of the minimum phoneme error criterion is as follows: (1) FMPE(Λ)=Σz=1ZΣszi∈SPΛ(szi|Xz)Raw Acc(SzR,szi) {F_{MPE}}(\Lambda) = \Sigma _{z = 1}^Z{\Sigma _{{s_{zi}} \in S}}{P_\Lambda}\left({{s_{zi}}|{X_z}} \right){\rm{Raw}}{\kern 1pt} {\rm{Acc}}\left({{S_{zR}},{s_{zi}}} \right)

Here, X_z is the feature vector sequence of sentence z; X_zR is the correct recognition result corresponding to X_z [21, 22]; S_zi is one of the candidate sequences of sentence z on the word graph; S_zR can be a phoneme according to the modelling unit and recognition task, syllables, words and strings, etc; Ss is the set of all possible candidate word sequences generated by the speech recogniser; and P_Λ (s_zi | X_z) is the posterior probability of the candidate word sequence S_zi. RawAcc(S_zR, s_zi) represents the number of correct identifications minus the number of insertion, deletion and substitution errors. By solving the auxiliary function constructed by Eq. (1), the expression of the first-order statistics of the numerator and denominator terms can be obtained (2), as follows: (2) θjmnum(X)=Σz=1ZΣq∈sziΣt=sqeqγqjm(t)max(0,γqzMPE)X(t)θjmden(X)=Σz=1ZΣq∈sziΣt=sqeqγqjm(t)max(0,−γqzMPE)X(t)γqzMPE=γq(c(q)−cavgz) \matrix{{\theta _{jm}^{num}(X) = \Sigma _{z = 1}^Z{\Sigma _{q \in {s_{zi}}}}\Sigma _{t = {s_q}}^{{e_q}}{\gamma _{qjm}}(t)\max \left({0,\gamma _q^{zMPE}} \right)X(t)} \hfill \cr {\theta _{jm}^{den}(X) = \Sigma _{z = 1}^Z{\Sigma _{q \in {s_{zi}}}}\Sigma _{t = {s_q}}^{{e_q}}{\gamma _{qjm}}(t)\max \left({0, - \gamma _q^{zMPE}} \right)X(t)\gamma _q^{zMPE} = {\gamma _q}\left({c(q) - c_{avg}^z} \right)} \hfill \cr}

The designed computer speech evaluation system can be divided into two parts: forced matching and score mapping. Force matching is used to get the GOP scores of phonemes, and score mapping maps the GOP scores of phonemes to words and sentences. The final evaluation system outputs an objective evaluation score of 0–100 based on phonemes, words and sentences.

First, the preprocessed learners' English pronunciation is verified by speech segment, including vowel segment cutting, establishment of verification system and reliability of verification system. The Viterbi algorithm cuts and decodes the speech segments and then sends the information including the extraction of evaluation parameters, the regularisation of evaluation parameters, the parameter association process and evaluation mechanism to the core of the English pronunciation evaluation system, that is, the pronunciation evaluation module. The weight of the evaluation parameters in the English pronunciation evaluation to reflects the opinions of human experts on the quality of English sentences and the feedback results provided to the learners including the comprehensive score and the correction opinions after the comparison of the expert knowledge base [23, 24, 25].

The so-called voice segment verification is to generate judgement thresholds for different evaluation voices and judge the correctness of the evaluation voice content according to the thresholds. After the verification system receives the evaluation speech segment, it performs pattern matching on each vowel, and then according to the distance of the matching result, combined with the verification mechanism, the final confidence threshold is given.

In this study, the forced alignment method commonly used in the Viterbi algorithm is used to cut the speech segment into smallest vowel pronunciation segments as much as possible. In this case, after the evaluation speech is cut, if the evaluation content and the standard pronunciation content are highly familiar, the number of vowel segments generated after cutting will be close to or equal to the number of standard pronunciation vowel segments. If the evaluation content is preceded by The n vowel consonant segments are the same as the standard pronunciation, and the following vowel consonant segments are speech segments outside the standard pronunciation library, so the number of vowel consonants after forced alignment is about n. After the vowel segment is cut, the number of vowel and consonant segments that can be cut out in the system is 15. For some speech segments that are not cut out, this study sets the confidence level to 0 to enhance the reliability of speech segment verification and make the recognition rate of the system higher. The pronunciation content of the upper part is the same as the standard pronunciation, and the system can cut out complete vowel segments; the pronunciation content of the lower part is partly different from the standard pronunciation, and the system can only cut out the vowel segments in the standard pronunciation library but cannot cut it. Vowel segments are outside the library.

In the aforementioned evaluation parameter extraction, the difference between the number of vowel segments obtained by cutting and the number of standard sentences is used to represent the completeness, and forced alignment is used to obtain the time correlation (speech rate change) and HMM log-probability distance of each vowel [33]. The extraction of the pitch period generally adopts the autocorrelation method and the method based on short-time amplitude difference, but the effect is not ideal and the timeliness is poor. In this study, the pitch period is obtained according to the time domain characteristics of the speech signal, and the process is as follows:

Find the maximum point of signal energy in a frame and record its time domain position.

After obtaining the degree of difference of each vowel in the four scoring parameters through the evaluation mechanism, the average degree of difference of a sentence is calculated according to the proportion, as shown in the following Table 1:

The computer objective speech evaluation system introduced in this study is designed for foreign language teaching and examination evaluation projects, and the purpose is to give evaluation scores that are more relevant to the subjective evaluation of experts. This study explores the use of discriminatively trained acoustic models to give evaluation scores that are more correlated with subjective expert evaluations. The mathematical theory of hypothesis testing proves that the discriminatively trained acoustic model can reduce the ‘false acceptance’ errors in the speech pronunciation evaluation system based on the VITERBI forced matching algorithm, thereby improving the confidence of the objective evaluation score. The relevant experiments based on the speech database collected in the fourth-level oral examination show that the discriminative acoustic model can give higher evaluation confidence scores than the acoustic model obtained by traditional maximum likelihood training.