Author Contributions
Conceptualization, E.G.-M., I.R., J.N. and P.C.M.; methodology, E.G.-M. and I.R.; software, J.N. and M.T.; validation, M.T. and F.P.; formal analysis, M.T., R.S.-S. and G.K.; investigation, J.N., M.T. and N.A.; resources, I.R., J.N. and M.T.; data curation, J.N. and M.T.; writing—original draft preparation, F.P.; writing—review and editing, F.P.; visualization, M.T.; supervision, R.S.-S., G.K. and F.P.; project administration, E.G.-M., I.R. and F.P. All authors have read and agreed to the published version of the manuscript.
Figure 1.
Simplified representation of how Parkinson’s disease affects speech and laughter. Speech/laughter decision-making cortical areas activate the motor commands–execution circuit (arrow 1a) as well as the basal ganglia–thalamus circuit (arrow 1b), which modulates the activity of these commands (arrow 3). Motor commands–execution areas send their output (arrow 4) to the motor nuclei which control muscles that generate speech/laughter sounds (arrow 5). In green, excitatory neuronal activity; in red, inhibitory neuronal activity; in grey, activity of dopaminergic neurons. Intense color indicates high neuronal activity; light color indicates low neuronal activity. In healthy subjects (left scheme), SNc-produced dopamine excites striatum neurons that inhibit SNr-GP inhibitory neurons. Low inhibitory input to the thalamus (arrow 2) is the ideal condition for the correct modulation of the motor commands (arrow 3), as well as the coordination of the motor nuclei (arrow 4) and of the corresponding muscles (arrow 5). In Parkinson’s disease (right scheme), the reduced SNc dopamine slows down striatum neurons, increasing SNr-GP inhibitory output (arrow 2). The inhibited thalamus fails in the modulation of cortical nuclei (arrow 3), losing the coordination of the motor nuclei (arrow 4) and provoking motor disorders (arrow 5). SNc, substantia nigra compacta; SNr, substantia nigra reticulata; GP, globus pallidus.
Figure 1.
Simplified representation of how Parkinson’s disease affects speech and laughter. Speech/laughter decision-making cortical areas activate the motor commands–execution circuit (arrow 1a) as well as the basal ganglia–thalamus circuit (arrow 1b), which modulates the activity of these commands (arrow 3). Motor commands–execution areas send their output (arrow 4) to the motor nuclei which control muscles that generate speech/laughter sounds (arrow 5). In green, excitatory neuronal activity; in red, inhibitory neuronal activity; in grey, activity of dopaminergic neurons. Intense color indicates high neuronal activity; light color indicates low neuronal activity. In healthy subjects (left scheme), SNc-produced dopamine excites striatum neurons that inhibit SNr-GP inhibitory neurons. Low inhibitory input to the thalamus (arrow 2) is the ideal condition for the correct modulation of the motor commands (arrow 3), as well as the coordination of the motor nuclei (arrow 4) and of the corresponding muscles (arrow 5). In Parkinson’s disease (right scheme), the reduced SNc dopamine slows down striatum neurons, increasing SNr-GP inhibitory output (arrow 2). The inhibited thalamus fails in the modulation of cortical nuclei (arrow 3), losing the coordination of the motor nuclei (arrow 4) and provoking motor disorders (arrow 5). SNc, substantia nigra compacta; SNr, substantia nigra reticulata; GP, globus pallidus.
Figure 2.
Temporal representation of one of the signals used in the study, followed by the steps of the analysis pipeline. DFT, digital Fourier transform. “Filter Banks” include Mel, Human Factor, and Bark filters.
Figure 2.
Temporal representation of one of the signals used in the study, followed by the steps of the analysis pipeline. DFT, digital Fourier transform. “Filter Banks” include Mel, Human Factor, and Bark filters.
Figure 4.
(a) SVM performance as a function of kernel selection. (b) kNN performance as a function of k, k = 1,5. AR with 90% confidence interval in both cases (Blue, MFCC. Yellow, BFCC. Red, HFCC). (c) Graphic representation of the RF model. Training data set is split into N different data subsets that feeds into the N generated decision trees (N = 100 in our study). Decision is taken following the final prediction, obtained by majority voting of the N decision trees, weighting the models according to their performance.
Figure 4.
(a) SVM performance as a function of kernel selection. (b) kNN performance as a function of k, k = 1,5. AR with 90% confidence interval in both cases (Blue, MFCC. Yellow, BFCC. Red, HFCC). (c) Graphic representation of the RF model. Training data set is split into N different data subsets that feeds into the N generated decision trees (N = 100 in our study). Decision is taken following the final prediction, obtained by majority voting of the N decision trees, weighting the models according to their performance.
Figure 5.
Representation of the receiver operating characteristics (ROC) curve for the three cepstral coefficients (MFCC, HFCC and BFCC) with the best performances. ROC relates the true positive rate (TPR) with the false positive rate (FPR) and the area under the curve (AUC) that measures the degree of separability between the two classes. Blue line represents an SVM with a ν-polynomial kerne, BFCC filter bank. Red line corresponds to an RF, MFCC filter bank. Yellow line corresponds to a kNN, k = 5, MFCC filter bank. 10-fold validation.
Figure 5.
Representation of the receiver operating characteristics (ROC) curve for the three cepstral coefficients (MFCC, HFCC and BFCC) with the best performances. ROC relates the true positive rate (TPR) with the false positive rate (FPR) and the area under the curve (AUC) that measures the degree of separability between the two classes. Blue line represents an SVM with a ν-polynomial kerne, BFCC filter bank. Red line corresponds to an RF, MFCC filter bank. Yellow line corresponds to a kNN, k = 5, MFCC filter bank. 10-fold validation.
Table 1.
Central frequencies corresponding to each of the 26 filters for the three scales employed in this study: Mel, Human Factor, and Bark.
Table 1.
Central frequencies corresponding to each of the 26 filters for the three scales employed in this study: Mel, Human Factor, and Bark.
Filter Nr | Mel (MFCC) | Human Factor (HFCC) | Bark (BFCC) |
---|
1 | 62.50 | 31.25 | 62.50 |
2 | 156.25 | 125.00 | 156.25 |
3 | 218.75 | 187.50 | 218.75 |
4 | 312.50 | 281.25 | 312.50 |
5 | 406.25 | 375.00 | 375.00 |
6 | 531.25 | 468.75 | 468.75 |
7 | 656.25 | 593.75 | 562.50 |
8 | 781.25 | 718.75 | 656.25 |
9 | 937.50 | 843.75 | 750.00 |
10 | 1093.75 | 1000.00 | 875.00 |
11 | 1250.00 | 1156.25 | 1000.00 |
12 | 1437.50 | 1343.75 | 1156.25 |
13 | 1656.25 | 1531.25 | 1281.25 |
14 | 1875.00 | 1781.25 | 1468.75 |
15 | 2125.00 | 2000.00 | 1656.25 |
16 | 2406.25 | 2281.25 | 1843.75 |
17 | 2718.75 | 2562.50 | 2093.75 |
18 | 3062.50 | 2875.00 | 2343.75 |
19 | 3437.50 | 3250.00 | 2656.25 |
20 | 3812.50 | 3625.00 | 3000.00 |
21 | 4281.25 | 4031.25 | 3406.25 |
22 | 4750.00 | 4500.00 | 3875.00 |
23 | 5281.25 | 5031.25 | 4406.25 |
24 | 5875.00 | 5537.50 | 5093.75 |
25 | 6531.25 | 6187.50 | 5937.50 |
26 | 7218.75 | 6875.00 | 6906.25 |
Table 2.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by individually employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Table 2.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by individually employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Results by employing μ, STD, skewness and kurtosis of the coefficients |
---|
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(MFCC) | 72 | 0.72 | 0.28 | 0.68 | 0.32 | 0.69 | 0.71 | 0.722 |
STD(MFCC) | 68 | 0.67 | 0.33 | 0.69 | 0.31 | 0.68 | 0.67 | 0.695 |
skew(MFCC) | 59 | 0.58 | 0.42 | 0.61 | 0.39 | 0.6 | 0.59 | 0.615 |
kurt(MFCC) | 60 | 0.62 | 0.38 | 0.59 | 0.41 | 0.6 | 0.61 | 0.625 |
μ(HFCC) | 72 | 0.72 | 0.28 | 0.69 | 0.32 | 0.7 | 0.71 | 0.725 |
STD(HFCC) | 70 | 0.7 | 0.3 | 0.69 | 0.31 | 0.7 | 0.7 | 0.721 |
skew(HFCC) | 65 | 0.65 | 0.35 | 0.65 | 0.35 | 0.65 | 0.65 | 0.67 |
kurt(HFCC) | 70 | 0.71 | 0.29 | 0.68 | 0.32 | 0.69 | 0.7 | 0.715 |
μ(BFCC) | 73 | 0.72 | 0.28 | 0.7 | 0.3 | 0.71 | 0.71 | 0.733 |
STD(BFCC) | 70 | 0.7 | 0.3 | 0.69 | 0.31 | 0.69 | 0.69 | 0.712 |
skew(BFCC) | 57 | 0.57 | 0.43 | 0.58 | 0.42 | 0.57 | 0.57 | 0.599 |
kurt(BFCC) | 63 | 0.65 | 0.35 | 0.62 | 0.39 | 0.63 | 0.63 | 0.654 |
Results by employing μ, STD, skewness and kurtosis of the delta (Δ) of the coefficients |
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(Δ(MFCC)) | 67 | 0.69 | 0.31 | 0.65 | 0.35 | 0.66 | 0.68 | 0.692 |
STD(Δ(MFCC)) | 74 | 0.7 | 0.3 | 0.65 | 0.35 | 0.66 | 0.68 | 0.694 |
skew(Δ(MFCC)) | 64 | 0.65 | 0.35 | 0.64 | 0.36 | 0.64 | 0.65 | 0.665 |
kurt(Δ(MFCC)) | 62 | 0.64 | 0.36 | 0.6 | 0.4 | 0.62 | 0.63 | 0.645 |
μ(Δ(HFCC)) | 69 | 0.7 | 0.3 | 0.68 | 0.32 | 0.69 | 0.69 | 0.712 |
STD(Δ(HFCC)) | 70 | 0.68 | 0.32 | 0.67 | 0.33 | 0.67 | 0.68 | 0.695 |
skew(Δ(HFCC)) | 70 | 0.7 | 0.3 | 0.71 | 0.29 | 0.7 | 0.7 | 0.72 |
kurt(Δ(HFCC)) | 64 | 0.67 | 0.33 | 0.61 | 0.39 | 0.63 | 0.65 | 0.664 |
μ(Δ(BFCC)) | 63 | 0.65 | 0.35 | 0.62 | 0.38 | 0.63 | 0.64 | 0.657 |
STD(Δ(BFCC)) | 71 | 0.68 | 0.32 | 0.69 | 0.31 | 0.69 | 0.69 | 0.71 |
skew(Δ(BFCC)) | 68 | 0.69 | 0.31 | 0.67 | 0.33 | 0.68 | 0.68 | 0.701 |
kurt(Δ(BFCC)) | 63 | 0.66 | 0.35 | 0.6 | 0.4 | 0.62 | 0.64 | 0.656 |
Results by employing μ, STD, skewness and kurtosis of the delta-delta (ΔΔ) of the coefficients |
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(ΔΔ(MFCC)) | 69 | 0.72 | 0.28 | 0.65 | 0.35 | 0.68 | 0.7 | 0.712 |
STD(ΔΔ(MFCC)) | 71 | 0.78 | 0.22 | 0.75 | 0.25 | 0.71 | 0.72 | 0.735 |
skew(ΔΔ(MFCC)) | 61 | 0.8 | 0.2 | 0.77 | 0.23 | 0.61 | 0.61 | 0.634 |
kurt(ΔΔ(MFCC)) | 66 | 0.79 | 0.21 | 0.77 | 0.23 | 0.66 | 0.66 | 0.685 |
μ(ΔΔ(HFCC)) | 69 | 0.71 | 0.29 | 0.66 | 0.34 | 0.68 | 0.7 | 0.713 |
STD(ΔΔ(HFCC)) | 71 | 0.73 | 0.27 | 0.69 | 0.31 | 0.7 | 0.72 | 0.734 |
skew(ΔΔ(HFCC)) | 66 | 0.65 | 0.35 | 0.66 | 0.34 | 0.66 | 0.65 | 0.675 |
kurt(ΔΔ(HFCC)) | 61 | 0.63 | 0.37 | 0.6 | 0.4 | 0.61 | 0.62 | 0.635 |
μ(ΔΔ(BFCC)) | 63 | 0.65 | 0.35 | 0.62 | 0.38 | 0.63 | 0.64 | 0.655 |
STD(ΔΔ(BFCC)) | 73 | 0.74 | 0.26 | 0.73 | 0.27 | 0.73 | 0.73 | 0.754 |
skew(ΔΔ(BFCC)) | 70 | 0.7 | 0.3 | 0.69 | 0.31 | 0.69 | 0.7 | 0.715 |
kurt(ΔΔ(BFCC)) | 60 | 0.58 | 0.42 | 0.62 | 0.38 | 0.6 | 0.6 | 0.626 |
Table 3.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by incrementally employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Table 3.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by incrementally employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Results by employing μ, STD, skewness and kurtosis of the coefficients |
---|
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(MFCC) | 72 | 0.72 | 0.28 | 0.68 | 0.32 | 0.69 | 0.71 | 0.71 |
μ+STD(MFCC) | 74 | 0.75 | 0.25 | 0.73 | 0.27 | 0.73 | 0.74 | 0.75 |
μ+STD+skew(MFCC) | 75 | 0.76 | 0.24 | 0.74 | 0.26 | 0.74 | 0.75 | 0.76 |
μ+STD+skew+kurt(MFCC) | 76 | 0.77 | 0.23 | 0.76 | 0.24 | 0.76 | 0.77 | 0.78 |
μ(HFCC) | 72 | 0.72 | 0.28 | 0.69 | 0.31 | 0.70 | 0.71 | 0.72 |
μ+STD(HFCC) | 74 | 0.74 | 0.26 | 0.73 | 0.27 | 0.74 | 0.74 | 0.76 |
μ+STD+skew(HFCC) | 76 | 0.77 | 0.23 | 0.75 | 0.25 | 0.76 | 0.76 | 0.78 |
μ+STD+skew+kurt(HFCC) | 77 | 0.79 | 0.21 | 0.76 | 0.24 | 0.77 | 0.78 | 0.80 |
μ(BFCC) | 73 | 0.72 | 0.28 | 0.70 | 0.30 | 0.71 | 0.71 | 0.73 |
μ+STD(BFCC) | 74 | 0.75 | 0.25 | 0.73 | 0.27 | 0.73 | 0.74 | 0.76 |
μ+STD+skew(BFCC) | 75 | 0.76 | 0.24 | 0.74 | 0.26 | 0.75 | 0.75 | 0.77 |
μ+STD+skew+kurt(BFCC) | 76 | 0.77 | 0.23 | 0.75 | 0.25 | 0.76 | 0.76 | 0.79 |
Results by employing μ, STD, skewness and kurtosis of the delta (Δ) of the coefficients |
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(Δ(MFCC)) | 67 | 0.69 | 0.31 | 0.65 | 0.35 | 0.66 | 0.68 | 0.69 |
μ+STD(Δ(MFCC)) | 72 | 0.73 | 0.27 | 0.72 | 0.28 | 0.72 | 0.73 | 0.75 |
μ+STD+skew(Δ(MFCC)) | 73 | 0.75 | 0.25 | 0.72 | 0.28 | 0.73 | 0.74 | 0.76 |
μ+STD+skew+kurt(Δ(MFCC)) | 75 | 0.76 | 0.24 | 0.75 | 0.25 | 0.75 | 0.76 | 0.78 |
μ(Δ(HFCC)) | 69 | 0.7 | 0.3 | 0.68 | 0.32 | 0.69 | 0.69 | 0.71 |
μ+STD(Δ(HFCC)) | 72 | 0.71 | 0.29 | 0.72 | 0.28 | 0.72 | 0.71 | 0.74 |
μ+STD+skew(Δ(HFCC)) | 73 | 0.73 | 0.27 | 0.73 | 0.27 | 0.73 | 0.73 | 0.75 |
μ+STD+skew+kurt(Δ(HFCC)) | 76 | 0.76 | 0.24 | 0.76 | 0.24 | 0.76 | 0.76 | 0.78 |
μ(Δ(BFCC)) | 63 | 0.65 | 0.35 | 0.62 | 0.38 | 0.63 | 0.64 | 0.66 |
μ+STD(Δ(BFCC)) | 67 | 0.67 | 0.33 | 0.68 | 0.32 | 0.67 | 0.67 | 0.69 |
μ+STD+skew(Δ(BFCC)) | 69 | 0.69 | 0.31 | 0.70 | 0.30 | 0.69 | 0.69 | 0.71 |
μ+STD+skew+kurt(Δ(BFCC)) | 71 | 0.72 | 0.28 | 0.72 | 0.28 | 0.72 | 0.72 | 0.74 |
Results by employing μ, STD, skewness and kurtosis of the delta-delta (ΔΔ) of the coefficients |
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
μ(ΔΔ(MFCC)) | 69 | 0.72 | 0.28 | 0.65 | 0.35 | 0.68 | 0.70 | 0.71 |
μ+STD(ΔΔ(MFCC)) | 76 | 0.78 | 0.22 | 0.75 | 0.25 | 0.76 | 0.77 | 0.79 |
μ+STD+skew(ΔΔ(MFCC)) | 78 | 0.79 | 0.21 | 0.77 | 0.23 | 0.78 | 0.79 | 0.81 |
μ+STD+skew+kurt(ΔΔ(MFCC)) | 78 | 0.80 | 0.20 | 0.77 | 0.23 | 0.78 | 0.79 | 0.81 |
μ(ΔΔ(HFCC)) | 69 | 0.71 | 0.29 | 0.66 | 0.34 | 0.68 | 0.70 | 0.71 |
μ+STD(ΔΔ(HFCC)) | 75 | 0.76 | 0.24 | 0.73 | 0.27 | 0.74 | 0.75 | 0.77 |
μ+STD+skew(ΔΔ(HFCC)) | 75 | 0.77 | 0.24 | 0.74 | 0.26 | 0.75 | 0.76 | 0.78 |
μ+STD+skew+kurt(ΔΔ(HFCC)) | 76 | 0.77 | 0.23 | 0.75 | 0.25 | 0.75 | 0.77 | 0.78 |
μ(ΔΔ(BFCC)) | 63 | 0.65 | 0.35 | 0.62 | 0.38 | 0.63 | 0.64 | 0.66 |
μ+STD(ΔΔ(BFCC)) | 72 | 0.73 | 0.27 | 0.72 | 0.28 | 0.72 | 0.72 | 0.74 |
μ+STD+skew(ΔΔ(BFCC)) | 73 | 0.73 | 0.27 | 0.73 | 0.27 | 0.73 | 0.73 | 0.75 |
μ+STD+skew+kurt(ΔΔ(BFCC)) | 74 | 0.75 | 0.26 | 0.74 | 0.26 | 0.74 | 0.74 | 0.76 |
Table 4.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by incrementally employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), together with their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Table 4.
Evaluation of the RF model with MFCC, HFCC and BFCC filters, by incrementally employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), together with their Δ and their ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Inputs | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
---|
μ(MFCC+Δ(MFCC)+ΔΔ(MFCC)) | 74 | 0.77 | 0.23 | 0.71 | 0.29 | 0.73 | 0.76 | 0.75 |
μ+STD(MFCC+Δ(MFCC)+ΔΔ(MFCC)) | 82 | 0.83 | 0.17 | 0.82 | 0.18 | 0.82 | 0.83 | 0.84 |
μ+STD+skew(MFCC+Δ(MFCC)+ΔΔ(MFCC)) | 83 | 0.84 | 0.16 | 0.82 | 0.18 | 0.82 | 0.84 | 0.85 |
μ+STD+skew+kurt(MFCC+Δ(MFCC)+ΔΔ(MFCC)) | 83 | 0.84 | 0.16 | 0.82 | 0.18 | 0.83 | 0.84 | 0.86 |
μ(HFCC+Δ(HFCC)+ΔΔ(HFCC)) | 75 | 0.77 | 0.23 | 0.73 | 0.27 | 0.74 | 0.76 | 0.76 |
μ+STD(HFCC+Δ(HFCC)+ΔΔ(HFCC)) | 81 | 0.82 | 0.18 | 0.81 | 0.19 | 0.81 | 0.82 | 0.83 |
μ+STD+skew(HFCC+Δ(HFCC)+ΔΔ(HFCC)) | 82 | 0.83 | 0.17 | 0.82 | 0.18 | 0.82 | 0.82 | 0.84 |
μ+STD+skew+kurt(HFCC+Δ(HFCC)+ΔΔ(HFCC)) | 82 | 0.83 | 0.17 | 0.82 | 0.18 | 0.82 | 0.83 | 0.85 |
μ(BFCC+Δ(BFCC)+ΔΔ(BFCC)) | 72 | 0.74 | 0.26 | 0.70 | 0.30 | 0.74 | 0.71 | 0.76 |
μ+STD(BFCC+Δ(BFCC)+ΔΔ(BFCC)) | 80 | 0.80 | 0.20 | 0.80 | 0.20 | 0.80 | 0.80 | 0.82 |
μ+STD+skew(BFCC+Δ(BFCC)+ΔΔ(BFCC)) | 81 | 0.81 | 0.19 | 0.81 | 0.19 | 0.81 | 0.81 | 0.84 |
μ+STD+skew+kurt(BFCC+Δ(BFCC)+ΔΔ(BFCC)) | 82 | 0.82 | 0.18 | 0.81 | 0.19 | 0.82 | 0.81 | 0.85 |
Table 5.
Results of the variation of the kernel in the SVM model with MFCC, HFCC and BFCC filters, by employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt). 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Table 5.
Results of the variation of the kernel in the SVM model with MFCC, HFCC and BFCC filters, by employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt). 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Best performance per column is highlighted in bold.
Results of Mel filters: μ + STD + skew + kurt (MFCC + Δ(MFCC) + ΔΔ(MFCC)) |
---|
Kernel variation | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
Linear | 74 | 0.74 | 0.26 | 0.73 | 0.27 | 0.73 | 0.74 | 0.76 |
Polynomial | 73 | 0.75 | 0.25 | 0.72 | 0.28 | 0.73 | 0.74 | 0.76 |
Radial Basis | 65 | 0.86 | 0.14 | 0.45 | 0.55 | 0.61 | 0.76 | 0.72 |
ν-Linear | 81 | 0.81 | 0.19 | 0.81 | 0.19 | 0.81 | 0.81 | 0.85 |
ν-Polynomial | 82 | 0.82 | 0.18 | 0.83 | 0.17 | 0.82 | 0.82 | 0.86 |
ν-Radial Basis | 73 | 0.85 | 0.15 | 0.60 | 0.40 | 0.68 | 0.80 | 0.79 |
Results of Human Factor filters: μ + STD + skew + kurt (HFCC + Δ(HFCC) + ΔΔ(HFCC)) |
Kernel variation | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
Linear | 74 | 0.74 | 0.26 | 0.73 | 0.27 | 0.74 | 0.74 | 0.78 |
Polynomial | 74 | 0.75 | 0.25 | 0.73 | 0.27 | 0.73 | 0.74 | 0.78 |
Radial Basis | 66 | 0.86 | 0.14 | 0.45 | 0.55 | 0.61 | 0.76 | 0.73 |
ν-Linear | 81 | 0.81 | 0.19 | 0.81 | 0.19 | 0.81 | 0.81 | 0.85 |
ν-Polynomial | 83 | 0.83 | 0.17 | 0.83 | 0.17 | 0.83 | 0.83 | 0.86 |
ν-Radial Basis | 73 | 0.85 | 0.15 | 0.61 | 0.39 | 0.69 | 0.81 | 0.79 |
Results of Bark filters: μ + STD + skew + kurt (BFCC + Δ(BFCC) + ΔΔ(BFCC)) |
Kernel variation | AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
Linear | 71 | 0.71 | 0.29 | 0.72 | 0.28 | 0.72 | 0.71 | 0.76 |
Polynomial | 72 | 0.72 | 0.28 | 0.72 | 0.28 | 0.72 | 0.72 | 0.76 |
Radial Basis | 63 | 0.85 | 0.15 | 0.41 | 0.59 | 0.59 | 0.73 | 0.69 |
ν-Linear | 80 | 0.80 | 0.20 | 0.80 | 0.20 | 0.80 | 0.80 | 0.85 |
ν-Polynomial | 82 | 0.82 | 0.18 | 0.82 | 0.18 | 0.82 | 0.82 | 0.86 |
ν-Radial Basis | 66 | 0.85 | 0.15 | 0.47 | 0.53 | 0.62 | 0.76 | 0.72 |
Table 6.
Summary of the results of the RF model with MFCC, HFCC and BFCC filters, by employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), Δ and ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Note that the three rows correspond to the 4th, 8th and 12th row of
Table 4.
Table 6.
Summary of the results of the RF model with MFCC, HFCC and BFCC filters, by employing the first four moments of their distributions (mean-μ, standard deviation-STD, skewness-skew and kurtosis-kurt), Δ and ΔΔ. 10-cross-validation with 20,000 laughs (18,000 training and 2000 test in 10 epochs). AR, accuracy rate; TP, true positive; FP, false positive; TN, true negative; FN, false negative; Sens, sensitivity; Spec, specificity. Note that the three rows correspond to the 4th, 8th and 12th row of
Table 4.
| AR (%) | TP | FP | TN | FN | Sens | Spec | AUC |
---|
MFCC | 83 | 0.84 | 0.16 | 0.82 | 0.18 | 0.83 | 0.84 | 0.86 |
HFCC | 82 | 0.83 | 0.17 | 0.82 | 0.18 | 0.82 | 0.83 | 0.85 |
BFCC | 81 | 0.82 | 0.18 | 0.81 | 0.19 | 0.82 | 0.81 | 0.85 |