**4. Feature analysis: acoustic analysis**

This part of the chapter presents a method, called feature analysis, based on the analysis of acoustic speech features. Children with SLI have a specific problem with the production and perception of spoken language as well as show signs of motor, auditory and phonology difficulties.

Achievements in the recognition of emotions were the inspiration for the use of this method of speech processing. An analogy between the research of emotions and the research of pathological speech, that is, speech of children with SLI, can be made. Typical children, who lack pathological changes in speech and a diagnosis of any disease, can be compared with a neutral emotion. Children with SLI can be compared with some unspecified emotion, for example, anger or fear.

This method is focused on acoustic speech parameters from individual words. Analysis was performed by comparing the words pronounced by the cases versus the words pronounced by the controls. The aim is to identify features that can be used to uniquely identify cases.

#### **4.1. Description**

Classification labeled as misclassification (*"misclass"* in **Figure 2**) indicates the values located outside these limits. As the final criterion for the classification of several words containing an error, the group of cases comprised all children who had more than six words with any error during testing. Self-Organizing Maps (SOMs) [30], a subgroup of an Artificial Neural Network (ANN) [31], was the basis for the other three methods. Parameters for the ANN were set with a standard approach, that is, ratios were 0.7 for training, 0.15 for testing and 0.15 for validation. Differences were observed in the values for weights. ANN1 comprised original default values, ANN2 comprised minimum and maximum values from both groups and ANN3 comprised the weights set to the mean values of these groups. **Figure 2** provides a

**Table 3.** Final results for classifiers based on the HM and ANN methods. The method with the highest rate of success

**95.35** 81.40 97.67 97.67

**92.60** 88.89 83.33 87.04

**Methods HM (%) ANN1 (%) ANN2 (%) ANN3 (%)**

All 38.89 4.93 33.96 2 vs. 1 688.84

**Table 2.** Error analysis: comparison of both groups: average number of errors for controls and for cases.

**Age category Average error Difference 2 vs. 1 Comparison Difference [%]**

**Final classification 93.81** 85.14 90.50 92.36

Approximately the same final results are observed for both classifiers, but the HM classifier is easier to implement and use. The results indicate that children with SLI had a greater number

**Figure 2.** Process diagram illustrating the principle of error analysis. Overview and comparison of the classification

process diagram for the classification of the error analysis method.

of errors in their utterances than typical children.

through ANN and HM method.

**Error analysis: the success of classification**

**Error analysis: controls vs. cases: participants**

10 Learning Disabilities - An International Perspective

**Cases (2) Controls (1)**

**Classification for controls (***p\_h***)**

**Classification for cases (***p\_sli***)**

in bold.

The examined issue, the classification of children with SLI, has the following implementation structure. The implementation can be divided into four parts whose respective key components can be described as follows:


(d) Classification: The classification of the children into two groups, controls and cases, was relatively simple. The evaluation of the selected speech features of all words (selected words listed in **Table 1**) in this study was performed. The resulting classification for each participant was evaluated from the winning class based on the number of classifications (i.e., a value of one corresponded to the correct classification, and a value of two corresponded to a misclassification). Two different approaches of feature selection were used (FS: constant and FS: variable) for classification. For the variable type, properties with the highest accuracy rate (with more than a 90% success of classification for each word) were selected. For the constant type of feature selection, participants were classified using 268 features obtained from 38 words. If a constant number of features (with the 30 best parameters for each word) was used, each participant was classified using 760 features. The entire process of identifying the children with SLI is shown in **Figure 4**. Comparison of the approaches of feature selection is shown in **Figure 5**. The success rate of classification of the FS variable is shown in red (or is presented as the top values), and that of the FS constant is shown in blue (or is presented as the bottom values) or by grayscale. The horizontal line in this chart is the critical line for classification success. The x-axis represents all words, and the y-axis represents the success rate as a percentage.

#### **4.2. Statistical evaluation and results**

The data were divided into four groups depending on the classification, that is, correct or incorrect classification for controls (*p\_h*) and correct or incorrect classification for cases (*p\_ sli*). The number of classifications was based on the evaluation of features. Statistical tests evaluated the correct versus incorrect classification of selected features for both groups of children.

The scores of the Shapiro-Wilks test for normality are as follows: for *p\_h*, correct: *W* = 0.5969 and *p-val* = 8.965e−10 and wrong: *W* = 0.5678 and *p-val* = 3.567e−10; for *p\_sli*, correct: *W* = 0.7825 and *p-val* = 1.598e−07 and wrong: *W* = 0.7898 and *p-val* = 2.344e−07. These values were too small (*p-val* < 0.05) to use to confirm the hypothesis that the groups have a normal distribution. The

**Figure 3.** Procedure for selecting appropriate acoustic features. The value of 1 indicates correct classification, and the value of 2 indicates incorrect classification. Parameters *"maxg1"* and "ming2" indicate maximum and minimum threshold limits for incorrect classification.

(d) Classification: The classification of the children into two groups, controls and cases, was relatively simple. The evaluation of the selected speech features of all words (selected words listed in **Table 1**) in this study was performed. The resulting classification for each participant was evaluated from the winning class based on the number of classifications (i.e., a value of one corresponded to the correct classification, and a value of two corresponded to a misclassification). Two different approaches of feature selection were used (FS: constant and FS: variable) for classification. For the variable type, properties with the highest accuracy rate (with more than a 90% success of classification for each word) were selected. For the constant type of feature selection, participants were classified using 268 features obtained from 38 words. If a constant number of features (with the 30 best parameters for each word) was used, each participant was classified using 760 features. The entire process of identifying the children with SLI is shown in **Figure 4**. Comparison of the approaches of feature selection is shown in **Figure 5**. The success rate of classification of the FS variable is shown in red (or is presented as the top values), and that of the FS constant is shown in blue (or is presented as the bottom values) or by grayscale. The horizontal line in this chart is the critical line for classification success. The x-axis represents all words, and the y-axis represents the success rate

The data were divided into four groups depending on the classification, that is, correct or incorrect classification for controls (*p\_h*) and correct or incorrect classification for cases (*p\_ sli*). The number of classifications was based on the evaluation of features. Statistical tests evaluated the correct versus incorrect classification of selected features for both groups of

The scores of the Shapiro-Wilks test for normality are as follows: for *p\_h*, correct: *W* = 0.5969 and *p-val* = 8.965e−10 and wrong: *W* = 0.5678 and *p-val* = 3.567e−10; for *p\_sli*, correct: *W* = 0.7825 and *p-val* = 1.598e−07 and wrong: *W* = 0.7898 and *p-val* = 2.344e−07. These values were too small (*p-val* < 0.05) to use to confirm the hypothesis that the groups have a normal distribution. The

**Figure 3.** Procedure for selecting appropriate acoustic features. The value of 1 indicates correct classification, and the value of 2 indicates incorrect classification. Parameters *"maxg1"* and "ming2" indicate maximum and minimum threshold

as a percentage.

limits for incorrect classification.

children.

**4.2. Statistical evaluation and results**

12 Learning Disabilities - An International Perspective

**Figure 4.** Process diagram illustrating the principle of feature analysis. Overview of the classification individual groups.

**Figure 5.** Feature analysis: improving of feature selection method. The success rate of classification of the FS variable is shown in red (or is presented as the top values), and that of the FS constant is shown in blue (or is presented as the bottom values) or by grayscale.

scores of the Wilcoxon rank-sum test, which was used as a substitute for the *t*-test, are as follows: for *p\_h*, correct vs. wrong: *p-val* = 1.7510e−15, *zval* = 7.9578 and *ranksum* = 2911; for *p\_sli*, correct vs. wrong: *p-val* = 3.3145e−19, *zval* = −8.9577 and *ranksum* = 1485. The null hypothesis of equal medians was rejected because the p-values were too small, that is, a smaller one than the significance level was set, and the values for the group were not the same at this significance level. These results indicate significant differences in the number of classifications between wrong and correct evaluations for controls and cases.

**Table 4** presents the final evaluation used to distinguish the two groups, that is, controls vs. cases. The success rate was almost 97%, exactly 96.94%. Three participants (from controls) out of 98 were classified incorrect. Obtained results proved that it is possible to find method based on the acoustic features that can distinguish typically children from children with SLI with high accuracy.

The results of the feature analyses for all participants are displayed in **Figure 6**. Correct classifications of the control group are displayed in blue (or at a higher position), and incorrect classifications of the cases are displayed in red (or at a lower position) or by grayscale. The upper graph showed the total number of classifications where the values in the higher positions indicate a more successful classification. The middle histogram represents the


**Table 4.** Evaluation of percent success rate of method based on the acoustic features. The final success rate is in bold.

**Figure 6.** Evaluation of the feature analysis for cases. The correct classification is shown in blue (or at a higher position), and incorrect classification is shown in red (or at a lower position) or in grayscale. Samples with a more successful classification are at a higher position in the upper graph. The histogram represents the distributions of the correct classifications of controls and incorrect classifications of cases (more classifications in the right) in the middle graph. The boxplots show significant differences between the correct (the left boxplot) and incorrect classifications (the right boxplot) in the bottom chart.

distributions of the correct classifications of controls and incorrect classifications of cases. Participants in the higher positions (in the right part of the chart) have more successful classifications. The bottom boxplots show significant differences between the correct (blue or the left boxplot) and incorrect classifications (red or the right boxplot). There was an analogous situation for cases.
