IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 1, March 2024, pp. 185~194

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp185-194  185

Effective modelling of human expressive states from voice by

adaptively tuning the neuro-fuzzy inference system

Surjyo Narayana Panigrahi1, Niharika Pattanaik2, Hemanta Kumar Palo2

1
Department of Electronics and Communication Engineering, YBN University, Jharkhand, India
2
Department of Electronics and Communication Engineering, Institute of Technical Education and Research,
Siksha O Anusandhan (Deemed to be University), Bhubaneswar, India

Article Info ABSTRACT

Article history: This paper aims to develop efficient speech-expressive models using the
adaptively tuning neuro-fuzzy inference system (ANFIS). The developed
Received Jul 30, 2022 models differentiate a high-arousal happiness state from a low-arousal sadness
Revised Feb 12, 2023 state from the benchmark Berlin (EMODB) database. The proposed low-cost
Accepted Mar 10, 2023 flexible developed algorithms are self-tunable and can address several vivid
real-world issues such as home tutoring, banking, and finance sectors,
criminal investigations, psychological studies, call centers, cognitive and
Keywords: biomedical sciences. The work develops the proposed structures by
formulating several novel feature vectors comprising both time and frequency
Adaptive neuro-fuzzy inference information. The features considered are pitch (F0), the standard deviation of
Feature extraction pitch (SDF0), autocorrelation coefficient (AC), log-energy (E), jitter,
Human expressive states shimmer, harmonic to noise ratio (HNR), spectral centroid (SC), spectral roll-
Modeling off (SR), spectral flux (SF), and zero-crossing rate (ZCR). to alleviate the
Root mean square error issues of the curse of dimensionality associated with the frame-level
extraction, the features are extracted at the utterance level. Several
performance parameters have been computed to validate the individual time
and frequency models. Further, the ANFIS models are tested for their efficacy
in a combinational platform. The chosen features are complementary and the
augmented vectors have indeed shown improved performance with more
available information as revealed by our results.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Hemanta Kumar Palo
Department of Electronics and Communication Engineering
Institute of Technical Education and Research Siksha O Anusandhan (Deemed to be University)
Bhubaneswar, Odisha, India
Email: [email protected]

1. INTRODUCTION
Human expressive states are highly unpredictable, vague, overlapping, and ill-defined. Human being
via facial expressions, gestures, and through voice modalities often manifests these states. Trivial things, such
as watching a movie, listening to songs, meeting an old-time friend, a pleasant scent, and seeing a funeral pyre
can change our expressive states. The study remains a complex domain of research, particularly when these
states are expressed via phone [1]. It requires effective signal-processing tools to adequately represent them
that can benefit several fields such as artificial intelligence, cognitive sciences, psychological studies, criminal
investigation and humanoid robotics. The tools and techniques must be capable of extracting discriminant and
relevant voice parameters appropriately representing human expressive states for efficient modeling [2].
Among several techniques, the community often relies on the prominent features extracted either at
the frame level or at the utterance level. The frame-level analysis facilitates studying the signal in a stationary

Journal homepage: http://ijai.iaescore.com

186  ISSN: 2252-8938

platform, however, results in high-dimensional data comprising redundant information, thus increasing the
training time and memory space [3], [4]. The utterance level extraction of the speech features can alleviate
these issues with improved accuracy [5]. Modeling algorithms often play a crucial role and remain an
indispensable module to develop an effective recognition system. Earlier Learning algorithms applied in this
field belong to neural networks, structural techniques, clustering approaches, Hidden Markov model, Gaussian
mixture model and support vector machines (SVM). with excellent outcomes [6]–[13]. The spectrograms of
speech emotions along with the squeeze and excitation residual neural network (ResNet) and a trainable
discriminative ghost vector of locally aggregated descriptors (GhostVLAD) clustering layer extended
convolution neural network (CNN) have been applied to extract a low-dimensional utterance-level feature
vector [5]. Simulation on crowd sourced emotional multimodal actors datase (CREMA-D), Ryerson audio-
visual database of emotional speech, and song (RAVDESS) using the developed model has provided a global
accuracy of 83.35% and 64.92% respectively. The combination of empirical mode decomposition and the
Teager-Kaiser Energy Operator time-frequency and cepstral features has provided improved expressive state
models than stand-alone feature vectors. The authors have reported an accuracy of 91.16% and 86.22%
respectively using the recurrent neural network (RNN) and SVM in modelling the chosen speech expressive
states [6]. The hybridization of the prosodic, cepstral, spectrum, and wavelet features has enhanced the
modelling capability of SVM in recognizing Arabic expressive states [7]. These pieces of literature reveal that
the hybridization of features and their effective selection often lead to enhanced models due to more available
information, however not without limitations [1]–[4], [6]–[13]. However, the judicious selection of relevant
features bearing complementary information challenges the community, hence motivating the authors.
Several issues that are inherently associated with hybrid models are the requirement of a large pool of
samples, response time, selection of hyper-parameters, kernel function, the number of hidden layers, nodes,
feature dimension, addressing the non-linear relationship among extracted parameters, regularization,
generalization and issues of overfitting [12]. Henceforth, the application of fuzzy-based approaches along with
global statistics seems a novel ideal to explore in real-world fuzzy environments. These models remain flexible,
providing several feasible solutions besides performing in dynamic, unpredicted, and vague environments.
Unlike other neural networks, the neuro-fuzzy inference systems when adaptively tuned lead to an outcome-
based adaptive neuro-fuzzy inference system (ANFIS) structure that does not require frame length
normalization. The application of state transition probability in ANFIS facilitates the representation of temporal
dynamics associated with the baseline parameters. The structure rapidly learns from the experimental data with
precision and certainty. The use of approximation while generalizing the network and lower errors than the
conventional NNs during memorization makes it versatile [14]–[16]. User transparency, adaptability to
nonlinear signals, faster training without expert knowledge, and representability of numerical and linguistic
information make the algorithm superior [17], thus providing the desired platform for this work.
In this paper, the authors attempt to model a few chosen speech expressive states using effective
features in section 2 and the ANFIS in section 3. The model can arguably limit the aforementioned issues by
representing the expressive state using both numerical and linguistic knowledge. It makes the model more
transparent and user-friendly due to low memorization errors. Finally, the adaptation capability, faster learning,
and nonlinear ability, of the developed model add value to the recognition mechanism. Section 4 validates and
discusses the developed models using the derived feature vectors considering three proposed instances whereas
section 5 concludes the work with a few possible future directions.

2. THE PROPOSED APPROACH

The Berlin emotional speech database (EMODB) dataset chosen in this work comprises seven
expressive states such as anger, happiness, boredom, anxiety, sadness, disgust, and neutral sampled at a rate of
48 kHz and are down-sampled to 16 kHz for convenience. It has been a widely accessed database used to
analyze speech emotion (SE) states, which makes the comparing platform uniform [18]–[22]. From this
database, this work compares the three expressive models based on the level of arousal. These states are
happiness (high-arousal), sadness (low-arousal), and neutral. Forty-five utterances of each state are used to
extract the chosen feature vectors and for further processing. Initially, the ANFIS structure is developed so that
it can learn from the extracted input feature vector and compute the consequent parameters by estimating the
premise parameters using subtractive clustering. The hybrid learning algorithm is used to train the ANFIS
structure based on the premise parameters for 10 iterations. Finally, the developed structure is tested to validate
its performance. The proposed ANFIS combination Model, shown in Figure 1 concatenates both the time and
frequency-domain utterance-level statistical feature vectors to develop the desired ANFIS model for each
expressive state.

Int J Artif Intell, Vol. 13, No. 1, March 2024: 185-194

Int J Artif Intell ISSN: 2252-8938  187

Figure 1. The proposed ANFIS modelling

The utterance-level statistics extracted from the frame-level features of a signal are the mean,
range, standard deviation, skewness, and kurtosis. The proposed frequency-domain ANFIS model is shown in
Figure 2. It considers five feature vectors spectral rolloff (SR), spectral flux (SF), spectral centroid (SC),
fundamental frequency (F0), and standard deviation of F0 (SF0). Each speech sample corresponding to the
chosen expressive state is pre-emphasized, normalized, and mean subtracted to spectrally flatten and reduce
the finite precision effects [23]–[25]. The proposed time-domain model is shown in Figure 3. It considers six
feature vectors such as the normalized log-energy, zero crossing rate (ZCR), jitter, Shimmer, auto-correlation
coefficients (AC), and harmonic noise ratio (HNR). The necessary rule base is formed during the ANFIS
training and testing to fetch the desired output. The objective is to develop an expressive model that can easily
adapt to the multi-environment scenario.

Figure 2. The proposed frequency-domain ANFIS model

Effective modelling of human expressive states from voice by adaptively … (Surjyo Narayana Panigrahi)
188  ISSN: 2252-8938

Figure 3. The proposed time-domain ANFIS model

3. THE ANFIS ALGORITHM

The ANFIS integrates the adaptive neural network (ANN) and fuzzy inference system (FIS)
algorithms to determine the model parameters using fuzzy if-then rules and appropriate membership functions
(MF) [26], [27]. There are five layers in this structure comprising adaptive nodes in layer-1 (fuzzy layer) and
layer-4 (de-fuzzy layer) whereas layers-2 (product), 3 (normalized), and 5 (total output) have fixed nodes each
staging a particular function. The rule can be formed corresponding to each extracted feature 𝑥 as if 𝑥 (𝑢1 ) is
𝐺𝑖 , 𝑥 (𝑢2 ) is 𝐻𝑖 , and 𝑥 (𝑢𝑙 ) is 𝐼𝑖 , then 𝑅𝑢𝑙𝑒𝑠𝑖 = 𝑝𝑖 𝑥 (𝑢1 ) + 𝑞𝑖 𝑥 (𝑢2 ) + ⋯ + 𝑟𝑖 𝑥 (𝑢𝑙 ) + 𝑎𝑖 , where 𝑥 (𝑢1 ), 𝑥
(𝑢2 ), …, 𝑥 (𝑢𝑙 ) is the input features. The terms 𝐺𝑖 , 𝐻𝑖 , … represent the fuzzy sets, and the terms 𝑝𝑖 , 𝑞𝑖 , … are
the design parameters estimated while training the structure.
The output of layer-1 considering 𝑥, 𝑦 as inputs, and 𝑧 as output is given by 𝑂1,𝑖 = 𝜇𝐺𝑖 (. ), where
𝜇𝐺𝑖 (. ) is the MF representing the inputs 𝑥 or 𝑦 corresponding to 𝐺𝑖 . The MF assigns linguistic labels such as
low or high or medium to specify the feature values of an input vector to quantify 𝐺𝑖 . The popular bell-shaped
MF having values between zero and one is chosen here and is represented for input 𝑥 as
1
𝜇𝐺𝑖 (𝑥) = 1+|(𝑢−𝑟 )⁄𝑝 2𝑞𝑖 (1)
𝑖 1|

By varying the premise parameters 𝑝, 𝑞, and 𝑟, it is possible to accommodate several MFs representing
the fuzzy set. The layer-2 having fixed circle nodes multiplies the extracted input features of vectors 𝑥, 𝑦, ….
The layer-3 fixed circle nodes estimate the 𝑖 𝑡ℎ rule’s firing strength using the firing strength of all the rules and
provide an output 𝑂3,𝑖 with normalized firing strength. The weights of adaptive layer-4 square nodes are
estimated as linear functions with Sugeno inference coefficients 𝑚𝑖 , 𝑛𝑖 , and 𝑠𝑖 . The output of layer 2, 𝑂2,𝑖 and
the output of layer 3, 𝑂3,𝑖 with 𝑣𝑖 as the firing strength of the rule are given in (2) and (3) respectively whereas
the layer-4 provides the consequent parameters and its weighted output is described by (4). Similarly, the single
circle layer-5 node provides the overall or the estimated Sugeno FIS model output and is given by (5). In this,
the hybridized ANN and FIS compute the consequent parameters in the forward pass by propagating the
information up to the fourth layer and optimizing the parameters using a least square regression algorithm.
However, a gradient descent algorithm optimizes the parameters of the premises.

𝑂2,𝑖 = 𝜇𝐺𝑖 (𝑥) × 𝜇𝐻𝑖 (𝑦) × … = 𝑣𝑖 , 𝑖 = 1,2, 3, … (2)

𝑣𝑖
𝑂3,𝑖 = 𝑣̅ 𝑖 = 𝑣 , 𝑗 = 1,2, … (3)
1 +𝑣2 +⋯

Int J Artif Intell, Vol. 13, No. 1, March 2024: 185-194

Int J Artif Intell ISSN: 2252-8938  189

𝑂4,𝑖 = 𝑣̅ 𝑖 𝑓𝑖 = 𝑣̅ 𝑖 (𝑚𝑖 𝑤1 + 𝑛𝑖 𝑤2 + ⋯ + 𝑠𝑖 ) (4)

∑𝑖 𝑣𝑖 𝑓 𝑖
𝑂5,𝑖 = ∑𝑖 𝑣̅ 𝑖 𝑓𝑖 = (5)
∑𝑖 𝑣 𝑖

4. THE RESULTS AND DISCUSSION

The simulation results using the extracted time and frequency domain features with the ANFIS
structure is provided in this section to validate the proposed work. The work initially develops the time and
frequency domain ANFIS structures. Further, the root mean square error (RMSE) while developing the ANFIS
models for different states of emotions is graphically shown for comparison. Finally, the ANFIS model using
both the time and frequency domain features has been developed and the error has been computed to validate
the efficacy of the combined model.
Figure 4 provides the frequency domain ANFIS structure comprising five inputs such as SR, SF, SC,
fundamental frequency (F0), the standard deviation of F0 (SF0), and one of the desired states as the output.
The training rows constitute the desired input-output pair of an individual expressive state while developing
the desired model of that statement using a set of chosen feature vectors. A similar time-domain ANFIS
structure has been developed using six inputs such as log-energy, zero crossing rate (ZCR), jitter, Shimmer,
auto-correlation coefficients, and harmonic noise ratio in Figure 5. The frequency and time-domain rules can
be viewed from the ANFIS rule viewer for the chosen states. The Frequency-domain rule viewer in,

Figure 4. The frequency domain ANFIS structure

Figure 5. The time domain ANFIS structure

Effective modelling of human expressive states from voice by adaptively … (Surjyo Narayana Panigrahi)
190  ISSN: 2252-8938

Figure 6 comprises the input and output rules for the state of happiness. The rule viewers can be
developed similarly for the sadness and neutral states using the five frequency-domain inputs and six time-
domain inputs. The rule viewer helps to investigate the crisp value of each state based on the inputs. Similarly,
the time-domain rule viewer with six inputs is developed the intelligent model synthesizes all the crisp input
terms describing the chosen expressive state while approximating the decision-making process. The model
utilizes the pre-defined membership functions instead of the quantitative terms of the features to map the input
feature vectors to the chosen output shape for such a purpose.
Figure 7 graphically analyses the training RMSE corresponding to the Happiness states using
frequency-domain feature vectors. The RMSE is estimated using ten epochs and is the difference between the
training output, and the FIS output. At each training epoch, the minimization of the error takes place to develop
the desired ANFIS model. The training however stops when the network converges or uses a stopping criterion.
At each epoch, the error between the measured and modeled values is estimated and minimized until the
network converges. The RMSE compared to that of the low-arousal sadness and the neutral state using
frequency-domain feature vectors can be observed similarly. It is found to be 0.64837, 0.67807, and 0.68863
corresponding to happiness, sadness, and neutral states respectively. It shows the suitability of ANFIS in
modeling the high-arousal Happiness state as compared to the low-arousal Sadness state.

Figure 6. The frequency-domain rule viewer

Figure 7. The ANFIS error (RMSE) for happiness using frequency-domain feature vector

Int J Artif Intell, Vol. 13, No. 1, March 2024: 185-194

Int J Artif Intell ISSN: 2252-8938  191

Figure 8 graphically analyses the training RMSE corresponding to the Happiness states using time-
domain feature vectors. The RMSE compared to that of the low-arousal sadness and the neutral state using can
be observed similarly. It is found to be 1.02, 1.1522, and 1.5327 corresponding to happiness, sadness, and
neutral states respectively. Figure 9 graphically analyses the training RMSE corresponding to the Happiness
states using the combined time-frequency feature vectors. The RMSE is found to be 0.3868, 0.58327, and
0.7896 corresponding to happiness, sadness, and neutral states respectively. The frequency domain features
are more informative, hence providing better modeling with lower RMSE than time-domain models.
Nevertheless, the combinational model has outperformed the individual models due to more emotionally
relevant available information as observed in Figure 7 through Figure 9. Figure 10 provides the testing RMSE
for the Sadness state using frequency-domain feature vectors. It resolves the issues of overfitting by optimizing
the MFs. The testing error also cross-validates the generated ANFIS models by testing their generalization
ability at each epoch. It shows how effectively the ANFIS models behave to the extracted testing feature
vectors.

Figure 8. The ANFIS error (RMSE) for happiness using time-domain feature vector

Figure 9. The ANFIS error (RMSE) for happiness using time-frequency-domain feature vector

A comparison of ANFIS performances with frequency and time-domain feature vectors is shown in
Table 1. It shows that the checking and testing error is always higher than the training error, however, the

Effective modelling of human expressive states from voice by adaptively … (Surjyo Narayana Panigrahi)
192  ISSN: 2252-8938

difference is meager, and the chosen expressive states can be materialized without overfitting. The time and
frequency-domain ANFIS models are validated using several performance parameters including the RMSE at
the start, at convergence, training, testing, and checking. The models of each expressive state have been trailed
using four, eight, ten, fifteen, and twenty epochs to minimize the RMSE. With an increase in the number of
epochs, the time to train, check and test the network has increased exponentially. The training error is reduced
due to extensive learning; however, the testing error has increased as a trade-off due to overfitting with poor
network generalization. On the contrary, with a small number of epochs, underfitting occurs due to inadequate
learning. The network has provided the optimum performance with ten epochs, hence chosen here. Among the
default FIS hybrid and back-propagation learning algorithms, the hybridization of the back-propagation and
least square has witnessed the lowest RMSE in all the chosen cases, hence is considered. It has been observed
that the ANFIS models of different expressive states using time-domain feature vectors have experienced
higher RMSE in all the cases as compared to the frequency-domain feature vectors due to less relevant
information as revealed in Table 1.

Figure 10. The testing RMSE for the sadness state using frequency-domain feature vectors

Table 1. Comparison of ANFIS performance parameters using frequency and time-domain feature vectors
Performance Parameters Happiness Sadness Neutral
Frequency Time Frequency Time Frequency Time
E1 0.64837 1.0253 0.67807 0.78967 0.68663 1.5327
E2 0.68441 1.2464 0.99422 0.1.4351 0.74127 1.6922
E3 0.74964 2.4167 0.99233 2.0197 0.73034 3.2346
E4 0.678071 1.27306 0.686631 1.15217 0.648371 8.62407
E5 0.686626 2.3666 0.686626 2.53514 0.648367 2.86288
R 10 17 14 21 13 25
N 176 247 128 303 164 359
L1 84 119 60 147 78 175
L2 140 204 100 252 130 300
L=L1+L2 224 323 160 399 208 475
I2 2 4 3 6 3 7
Clustering algorithms Default Subtractive clustering parameters
− Range of Influence: 0.5
− Reject ratio: 0.15
− Squash factor: 1.25
Accept ratio: 0.5
M 3 (High, Low, and Medium)
I1 10
E1: Average FIS training output error, E2: Average FIS checking output error, E3: Average FIS testing output error, E4: ANFIS error
at start, E5: ANFIS error at the convergence, R: Number of rules, N: Number of nodes, L1: Number of linear parameters, L2: Number
of nonlinear parameters, M: Number of inputs MFs, I1: Number of epochs considered, I2: Number of epochs for convergence

Int J Artif Intell, Vol. 13, No. 1, March 2024: 185-194

Int J Artif Intell ISSN: 2252-8938  193

5. CONCLUSION
This piece of work attempts to investigate happiness, sadness, and neutral expressive states using an
efficient soft computing approach. In this process, the ANFIS algorithm has been explored to model the chosen
expressive states based on a few of the efficient time and frequency-domain utterance level features. Further,
the ANFIS models are validated in a time and frequency combinational platform for better efficacy. Several
performance parameters have been computed to test and check the developed models for their efficient
portrayal of expressive states. It can be inferred that the feature combination indeed provides improved models
due to the availability of more complementary information. It has witnessed the lowest training, testing, and
checking RMSE as compared to either the frequency or time-domain feature vectors. Investigation and
validation of other efficient feature extraction algorithms in combinational and reduction platforms may
provide new insights into this field.

REFERENCES
[1] Y. Cimtay, E. Ekmekcioglu, and S. Caglar-Ozhan, “Cross-subject multimodal emotion recognition based on hybrid fusion,” IEEE
Access, vol. 8, pp. 168865–168878, 2020, doi: 10.1109/ACCESS.2020.3023871.
[2] T. M. Wani, T. S. Gunawan, S. A. A. Qadri, M. Kartiwi, and E. Ambikairajah, “A comprehensive review of speech emotion
recognition systems,” IEEE Access, vol. 9, pp. 47795–47814, 2021, doi: 10.1109/ACCESS.2021.3068045.
[3] I. K. Fodor, “A survey of dimension reduction techniques,” Library, vol. 18, no. 1, pp. 1–18, 2002, doi: 10.2172/15002155.
[4] J. Yuan, L. Chen, T. Fan, and J. Jia, “Dimension reduction of speech emotion feature based on weighted linear discriminant
analysis,” International Journal of Signal Processing, Image Processing and Pattern Recognition, vol. 8, no. 11, pp. 299–308,
2015, doi: 10.14257/ijsip.2015.8.11.27.
[5] B. Mocanu, R. Tapu, and T. Zaharia, “Utterance level feature aggregation with deep metric learning for speech emotion
recognition,” Sensors, vol. 21, no. 12, 2021, doi: 10.3390/s21124233.
[6] L. Kerkeni, Y. Serrestou, K. Raoof, M. Mbarki, M. A. Mahjoub, and C. Cleder, “Automatic speech emotion recognition using an
optimal combination of features based on EMD-TKEO,” Speech Communication, vol. 114, pp. 22–35, 2019, doi:
10.1016/j.specom.2019.09.002.
[7] L. Abdel-Hamid, “Egyptian Arabic speech emotion recognition using prosodic, spectral and wavelet features,” Speech
Communication, vol. 122, pp. 19–30, 2020, doi: 10.1016/j.specom.2020.04.005.
[8] M. N. Mohanty and H. K. Palo, “Segment based emotion recognition using combined reduced features,” International Journal of
Speech Technology, vol. 22, no. 4, pp. 865–884, 2019, doi: 10.1007/s10772-019-09628-3.
[9] D. Li, Y. Zhou, Z. Wang, and D. Gao, “Exploiting the potentialities of features for speech emotion recognition,” Information
Sciences, vol. 548, pp. 328–343, 2021, doi: 10.1016/j.ins.2020.09.047.
[10] M. B. Akçay and K. Oğuz, “Speech emotion recognition: Emotional models, databases, features, preprocessing methods, supporting
modalities, and classifiers,” Speech Communication, vol. 116, pp. 56–76, 2020, doi: 10.1016/j.specom.2019.12.001.
[11] S. G. Koolagudi, Y. V. S. Murthy, and S. P. Bhaskar, “Choice of a classifier, based on properties of a dataset: case study-speech
emotion recognition,” International Journal of Speech Technology, vol. 21, no. 1, pp. 167–183, 2018, doi: 10.1007/s10772-018-
9495-8.
[12] M. Shah Fahad, A. Ranjan, J. Yadav, and A. Deepak, “A survey of speech emotion recognition in natural environment,” Digital
Signal Processing: A Review Journal, vol. 110, 2021, doi: 10.1016/j.dsp.2020.102951.
[13] L. Sun, S. Fu, and F. Wang, “Decision tree SVM model with Fisher feature selection for speech emotion recognition,” Eurasip
Journal on Audio, Speech, and Music Processing, vol. 2019, no. 1, 2019, doi: 10.1186/s13636-018-0145-5.
[14] S. Lalitha, D. Geyasruti, R. Narayanan, and M. Shravani, “Emotion detection using MFCC and cepstrum features,” Procedia
Computer Science, vol. 70, pp. 29–35, 2015, doi: 10.1016/j.procs.2015.10.020.
[15] L. Chen, W. Su, Y. Feng, M. Wu, J. She, and K. Hirota, “Two-layer fuzzy multiple random forest for speech emotion recognition
in human-robot interaction,” Information Sciences, vol. 509, pp. 150–163, 2020, doi: 10.1016/j.ins.2019.09.005.
[16] R. Ram, H. K. Palo, M. N. Mohanty, and L. P. Suresh, “Design of FIS-based model for emotional speech recognition,” Advances
in Intelligent Systems and Computing, vol. 397, pp. 77–88, 2016, doi: 10.1007/978-81-322-2671-0_8.
[17] M. Asimuzzaman, P. D. Nath, F. Hossain, A. Hossain, and R. M. Rahman, “Sentiment analysis of bangla microblogs using adaptive
neuro fuzzy system,” ICNC-FSKD 2017-13th International Conference on Natural Computation, Fuzzy Systems and Knowledge
Discovery, pp. 1631–1638, 2018, doi: 10.1109/FSKD.2017.8393010.
[18] L. Tan et al., “Speech emotion recognition enhanced traffic efficiency solution for autonomous vehicles in a 5G-enabled space-air-
ground integrated intelligent transportation system,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 3, pp.
2830–2842, 2022, doi: 10.1109/TITS.2021.3119921.
[19] S. Giripunje and N. Bawane, “ANFIS based emotions recognision in speech,” in Knowledge-Based Intelligent Information and
Engineering Systems, Berlin, Heidelberg: Springer Berlin Heidelberg, pp. 77–84. doi: 10.1007/978-3-540-74819-9_10.
[20] M. Viswanathan, Z. X. Zhang, X. W. Tian, and J. S. Lim, “Emotional-speech recognition using the neuro-fuzzy network,”
Proceedings of the 6th International Conference on Ubiquitous Information Management and Communication, ICUIMC’12, 2012,
doi: 10.1145/2184751.2184863.
[21] K. Wang, G. Su, L. Liu, and S. Wang, “Wavelet packet analysis for speaker-independent emotion recognition,” Neurocomputing,
vol. 398, pp. 257–264, 2020, doi: 10.1016/j.neucom.2020.02.085.
[22] H. Zhang, H. Huang, and H. Han, “A novel heterogeneous parallel convolution Bi-LSTM for speech emotion recognition,” Applied
Sciences (Switzerland), vol. 11, no. 21, 2021, doi: 10.3390/app11219897.
[23] M. Seo and M. Kim, “Fusing visual attention cnn and bag of visual words for cross-corpus speech emotion recognition,” Sensors
(Switzerland), vol. 20, no. 19, pp. 1–21, 2020, doi: 10.3390/s20195559.
[24] H. K. Palo and S. Sagar, “Comparison of neural network models for speech emotion recognition,” Proceedings-2nd International
Conference on Data Science and Business Analytics, ICDSBA 2018, pp. 127–131, 2018, doi: 10.1109/ICDSBA.2018.00030.
[25] K. Chauhan, K. K. Sharma, and T. Varma, “Speech emotion recognition using convolution neural networks,” Proceedings-
International Conference on Artificial Intelligence and Smart Systems, ICAIS 2021, pp. 1176–1181, 2021, doi:
10.1109/ICAIS50930.2021.9395844.

Effective modelling of human expressive states from voice by adaptively … (Surjyo Narayana Panigrahi)
194  ISSN: 2252-8938

[26] V. Rezaie, A. Parnianifard, D. Z. Rodriguez, S. Mumtaz, and L. Wuttisittikulkij, “Speech emotion recognition using ANFIS and
PSO-optimization with Word2Vec,” J Neuro Spine, vol. 1, no. 1, pp. 41–56, 2023, doi: 10.21203/rs.3.rs-1237929/v1.
[27] M. Dirik, “Optimized anfis model with hybrid metaheuristic algorithms for facial emotion recognition,” International Journal of
Fuzzy Systems, 2022, doi: 10.1007/s40815-022-01402-z.

BIOGRAPHIES OF AUTHORS

Surjyo Narayana Panigrahi is a Ph.D. research scholar at YBN University,

Namkum, Ranchi, Jharkhand, India. His area of interest belongs to Signal and Image. He is
involved in teaching and research for the last 18 years in different Engineering colleges in
Odisha, India. He can be contacted at email: [email protected].

Niharika Pattanaik is a Ph.D. research scholar at SOA University. Her area of

interest belongs to Signal and Image. He is involved in teaching and research for the last 9
years in different Engineering colleges in Odisha, India. She can be contacted at email:
[email protected].

Hemanta Kumar Palo received a Master of Engineering from Birla Institute of

Technology, Mesra, Ranchi in 2011 and a Ph.D. in 2018 from the Siksha 'O' Anusandhan
(Deemed to be University), Bhubaneswar, Odisha, India. Currently, he is serving as an
Associate Professor in the Department of Electronics and Communication Engineering at the
Institute of Technical Education and Research, Siksha 'O' Anusandhan University,
Bhubaneswar, Odisha, India. His area of research includes signal processing, speech and
emotion recognition, machine learning, and analysis of power quality disturbances. He can
be contacted at email: [email protected].

Int J Artif Intell, Vol. 13, No. 1, March 2024: 185-194