Computer English Speech Independent Evaluation System of the Fusion Discrimination Training Algorithm

— This paper proposes a discriminative training algorithm for autonomous speech evaluation of computer English. Firstly, the mathematical expression of discriminative training algorithm is defined, and the conditions of using the algorithm are deduced. To facilitate the discriminant training algo-rithm calculation, through the parameter usage and frequency to calculate the algorithm, thus simplifying the discriminative training algorithm for English speech evaluation method. Experimental results show that the proposed discriminative training algorithm is correct for computer English speech independent evaluation


Introduction
Computer English speech refers to the independent evaluation rules have been set on the basis of good, all kinds of interactive [1] and other subjects of environmental development, through the initiative to achieve the interaction of information feedback, in order to fully assess the results of computer English pronunciation [2], the need for independent assessments of the information obtained, to improve their English pronunciation the level of [3][4]. If there is a strong adaptability in a certain level of English pronunciation assessment, this adaptability can be transferred to other levels of voice evaluation, and ultimately make the whole English pronunciation assessment innovation ability can be improved [5]. At present, the computer English voice evaluation is a short time development area, and more attention is paid to the research and analysis of computer English voice problems based on the discriminative training algorithm [6].
Combined with the existing research results, this paper proposes a computer English speech autonomous evaluation system based on discriminative training algorithm under the condition of performance constraints. Moreover, the system can analyze the computer English speech by adjusting the speech frequency. Although in this paper, the design of the proposed discriminative training algorithm is not very complete, but from the point of mathematics, discriminative training proved to maximum conditions of use, but also the use of the experiment is validated and analyzed, these for future research is of significance.

Design of discriminative training algorithm
Based on the discriminative training algorithm, if the load unit used is unified, then the discriminative training rules can be recognized at this time, and then the effective analysis of the optimization effect is achieved. The reason why we need to define the discriminative training algorithm is to better evaluate and guide the optimization of discriminative training algorithms. And the definition of the load unit does not have much impact on this purpose, therefore, it is general. In the study of the =1. set, and in the specific evaluation process, can be combined with the actual situation of the discriminative training algorithm parameters are adjusted.
We replace FLOPS and Watt with task L (T) and English speech E (T) which are processed by system T in time T, we define ( ) T ! as: Discriminant training algorithm unit marked as ! , U Joule For the discriminative training algorithm, assume that there are N total computing nodes, which can be written as ( ) In the discriminative training algorithm, the corresponding rated frequency is set for each link. If the parameter runs at full load, then the ratio between the adjusted value of each link and the actual parameter can be regarded as the discriminative training theoretical value. However, the analysis of the specific operating environment, the algorithm features will have a greater impact on parameter usage, and if the workload changes, the distinction between training parameters will change, therefore, requires the actual evaluation of the value to calculate the distinction Training algorithm.

3
Distinctive training algorithm of English voice independent evaluation 3.1 Evaluation of discriminative training algorithms ( ) L T and ( ) E T evaluation as the main part of the evaluation, the general operating system, there is a corresponding distinction between the training parameters and the use of the interface in the model of each node, the parameters are set monitoring agent, the use of this distributed programming Environment, better able to synchronize these monitoring agents, but also the ability to quickly summarize the data. Taking f and ! to represent the training parameters and the usage rate respectively, the distinction between the two sampling interval are t ! time, then in the0-T moment, the total can be sampled for M times, combined with the above definition of the integral can be drawn: If the t ! tends to be infinitesimal, then by formula (5) calculated value that corresponds to the value of ( ) L T . In the formula, ( ) f t j ! " and ( ) t j ! " # are able to monitor and control agent evaluation, the agent will be in accordance with the t ! time interval set to continuous sampling, in order to obtain parameters and usage.
When the T time is over, the operation node evaluation values are aggregated, and the corresponding load of the whole model can be obtained according to formula (5).
For ( ) E T , it can be evaluated in a variety of ways. Considering that the computer has its rated operating frequency, its frequency level is dynamic in actual operation.
Therefore, it is very difficult to obtain the mathematical expression of ( ) i p t , and the formula (3) can be used to obtain the similar ( ) E T of English speech, and the realtime frequency data acquisition with t ! as the time interval is similar to the formula

Extreme value analysis of discriminative training algorithm
In the discriminated training analysis this paper, if in the time period, ( ) t ! is the maximum, then ( ) T ! is the largest, so: We first study the relationship between ( ) From Eq. (8), when ( ) t ! is fixed, there is an extremum of Eq. (9), and we can prove that this value is a maximum: Combined with the above discussion, we can see that if , then the corresponding discriminant training algorithm will be the largest. The results obtained in this paper are based on the deduction of mathematical theory, which will be verified by experiments.

Experiment analysis
The above empirical study shows that the application of discriminative training algorithm in computer English pronunciation is effective. It can evaluate the parameter usage rate ( ) Analysis of Figure 1 shows that the curves of both ( ) t ! and ( ) p t show similar changes, and the time points of reaching the peak and valley are basically the same, and the changes are convergent. Figure 2 on the right is an image obtained by enlarging Figure 1, and analyzing the two graphs reveals that, regardless of the set frequency level, the ( ) t ! curve and the independent evaluation curve obtained when the parameter work is analyzed are With the same trend of change, there is a linear relationship between them. In-depth analysis of its subtle changes, in Figure 2, a small number of ( ) p t extreme and ( ) t ! extreme, the point in time is inconsistent, there is a certain error, analysis of the reasons, mainly due to the evaluation frequency is inevitable Has a certain hysteresis, in Figure 2, most of the time with the frequency value are consistent, the entire curve also showed a trend of rising ladder. Mainly because, compared with the monitoring agent sampling frequency of use, evaluation instrument has a smaller frequency. In short, equation (7) is validated in our experiments. That is, if f is fixed, the existence of Then use the experiment 2 to analyze the parameters of the algorithm.  . GHz, 2.8GHz is the maximum parameter frequency set by the model in this paper. If the theoretical maximum value of the discriminative training algorithm changes, the theoretical value of the parameter frequency will be close to its maximum value. Therefore, in Figure 3, the entire curve is ascending, and the maximum of discriminative training also occurs at f= 2.8GHz , which is consistent with the theoretical derivation above. According to formula (7), we can get the formula of frequency: ( ) 3 3 Assuming that ! =1, can be concluded that the more peripheral power consumption is present, the faster the work is required in order to be able to achieve efficient and inefficient use of these extra English voices so that discriminative training algorithms can be improved. In the experiment designed in this paper, we did not introduce the actual parameters. If we take it into account, we can get , then max f =3.0353, which is greater than the maximum frequency of the parameter.

Conclusion
In this paper, a discriminative training algorithm is proposed for the computer English speech independent evaluation system, and the mathematic principle, concrete expression formulas and evaluation methods of the algorithm are described in detail, and the idea of further optimizing the discriminative training algorithm is put forward. The mathematic expression formula of the algorithm is obtained, and its maximum conditions of use are deduced, and the results are verified by experiments. At the same time, it is difficult to obtain the real-time frequency of speech in the evaluation of discriminative training algorithm. The model partition training algorithm proposed in this paper fully takes into account the optimization of the algorithm parameters, which can make up for the deficiencies in the previous formulas, and is able to independently evaluate these two students based on these two students at the same time. Finally, the correctness of the experiment is also verified.