A017410108

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 4, Ver. I (July – Aug. 2015), PP 01-08
www.iosrjournals.org
DOI: 10.9790/0661-17410108 www.iosrjournals.org 1 | Page
Emotion Recognition using combination of MFCC and LPCC
with Supply Vector Machine
Soma Bera1
, Shanthi Therese2
, Madhuri Gedam3
1
(Department of Computer Engineering, Shree L.R. Tiwari College of Engineering, India)
2
(Department of Information Technology, Thadomal Shahani Engineering College, India)
3
(Department of Computer Engineering, Shree L.R. Tiwari College of Engineering, India)
Abstract: Speech is a medium through which emotions are expressed by human being. In this paper, a mixture
of MFCC and LPCC has been proposed for audio feature extraction. One of the greatest advantage of MFCC is
that it is capable of identifying features even in the existence of noise and henceforth it is combined with the
advantage of LPCC which helps in extracting features in low acoustics. Two databases have been considered
namely Berlin Emotional Database and a SAVEE database. SVM has been implemented for classification of
seven emotions (Sadness, Joy, Fear, Anger, Boredom, Disgust and Neutral). Accuracy of the developed model
is presented using confusion matrix. The Recognition outcome of the combined MFCC and LPCC extracted
features are compared with the isolated results. The maximum accuracy rate that reaches by using the
combination of feature extraction method is 88.59% for non linear RBF (Radial Basis Function) kernel SVM.
Keywords: Confusion matrix, Feature extraction, Linear SVM, LPCC, MFCC, RBF non linear SVM, Speech
Emotion Recognition
I. Introduction
Speech contains emotions which reflects the mood of a person that describes the physical state of mind.
Human speech entails of various different emotions such as happiness, anger, sad, fear, disgust, boredom,
compassion, surprise and neutral. A vocal communication through which people express ideas has ample
amount of information that is understood discreetly. This information may be expressed in the intonation, pitch,
volume and speed of the voice as well as through emotional state of mind. The speaker‟s emotional state is
closely related to this information. The most common basic emotions are happiness (joy), anger, fear, boredom,
sadness, disgust and neutral. Over few decades, the recognition of emotions has become a vital research area
that has gained boundless interest.
II. Proposed Work
In our work, we combine the advantages of both the feature extraction techniques namely MFCC and
LPCC. The extracted audio features of both the techniques are then supplied to SVM classifier for final emotion
classification to reach an appropriate level of accuracy. The proposed method is based on the following four
principals [1].
Feature Extraction: In this, the speech signal is elaborated in order to obtain a definite number of
variables, called features, which resembles to be useful for speech emotion recognition.
Feature Selection: it chooses the most suitable features in order to reduce the computational load
thereby reducing the time required to identify an emotion.
Database: it is the memory of the classifier; it contains sentences (audio clips) separated according to
the emotions to be recognized.
Classification: this actually classifies the emotions by using the features selected by the Feature
Selection block and the audio clips in the Database.
III. Emotional Database
In current research work, an ample amount of research work has been carried out in the field of
Emotion Recognition. Various databases have been built for the same. In our case, we have used a Berlin
Emotional Database (BED), which is a German corpus and established by Department of acoustics technology
of Berlin Technical University [2]. Berlin Database plays a significant role in Speech Emotion Recognition. The
entire database is easily accessible and is freely available. Moreover, it is one of the most important database for
Emotion Recognition. Berlin Database consists of in total 800 utterances, out of which 535 utterances are
successfully extracted in this project. The database consists of 7 emotions namely Joy, Anger, Sadness, Fear,
Disgust, Boredom and Neutral. Another database that we are experimenting on is the one based on English
corpus known as Surrey Audio-Visual Expressed Emotion “SAVEE” database. This database consists of 7
emotions namely joy, Surprise, anger, sadness, fear, disgust, and neutral. There are four male speakers recorded

Emotion Recognition using combination of MFCC and LPCC with Supply Vector Machine
for the SAVEE database. The database consists of audio WAV files sampled at a frequency of 44.1 kHz. There
are 15 sentences for each of the seven emotion categories, 480 British English utterances in total [3].
IV. Feature Extraction
Various feature extraction techniques are available such as LPCC, MFCC, LPC, Perceptual Linear
Predictive Coefficient (PLP), Power Spectral Analysis (FFT), Mel Scale Cepstral Analysis (MEL), etc. We have
combined the advantages of MFCC and LPCC feature extraction in this project in order to reach a better
accuracy level for emotion recognition.
4.1 MFCC
The reason for MFCC being most widely used is its capability of easy calculation, noble ability of the
peculiarity, anti-noise and other various such advantages. MFCC in the low frequency region has a noble
resolution of frequency, and the robustness to noise is also very good, but the accuracy of the coefficient in the
high frequency is not satisfactory. The process of calculating MFCC is shown in Fig (1).
Fig (1) Process of Calculating MFCC
In the pre-processing stage first each signal is de-noised and are then divided into frames using a
Hamming window. In framing, the speech samples are segmented into the small frames with the time length
within the range of 20-40ms. In our case, we keep the frame length to be of 20ms. Next, in windowing phase,
each individual frame is windowed in order to minimize the signal discontinuities at the beginning and end of
each frame. The next phase known as FFT transforms each frame of N samples from the time domain into the
frequency domain. Mel Filter Bank consists of overlapping triangular filters in order to filter out any noise in
the speech sample. It is calculated by using the following equation (1) [4] [5]:
B(f) = 2595.log(1 + f/700) ………. (1)
where f resembles the frequency denoted on a linear scale axis.
Taking logarithm simply converts the multiplication of the magnitude in the Fourier transform into addition.
DCT converts the frequency domain back to time domain and finally the MFCC features of the speech samples
are extracted successfully.
4.2 LPCC
LPCC represents the characteristics of certain speech channel, and the same person with dissimilar
emotional speech will have dissimilar channel features, thereby extracting these feature coefficients to classify
the emotions contained in speech. The computational process of LPCC is usually a repetition of computing the
linear prediction coefficients (LPC) [6] means the process of calculating LPCC features remain the same as that
of calculating LPC features.
Fig (2) Process of calculating LPCC

The signal s[n] is predicted by a linear combination of its past values. The predictor equation (2) is defined as
………………… (2)
Here s[n] is the signal, ak is the predictor coefficients and s[n-k] is the predicted signal. The predicted error
signal in equation (3), is defined as
E[n] = s[n] - s[n] ………………… (3)
the linear prediction coefficients are transformed into cepstral coefficients using a recursive relation as
illustrated in equation (4) [8].
IDFT(log(|DFT(x)|)) ……………………. (4)
If x is LPC, the cepstral coefficients are known as linear prediction cepstral coefficients (LPCC).
V. SVM Classification
SVM, a binary classifier is a simple and effective computation of machine learning algorithms, and is
generally used for pattern identification and classification problems, and in case of restricted training data, it can
have a very noble classification performance as compared to other classifiers [7]. The different SVM that are
available for classification are linear and nonlinear SVM. Linear are the ones that classify the data elements
which have a linear hyperplane, meaning classes separated by a straight line. And the nonlinear are the ones
wherein the separating hyperplane are nonlinear that is not a straight line. The function of SVM is to transform
the original input set to a high dimensional feature space by making use of kernel function. Hence, nonlinear
problems can be solved by doing this transformation. Different SVM techniques are available such as One vs
One, One vs All, Polykernel SVM, Gaussian kernel SVM, Radial Basis Function (RBF), etc. RBF kernel is
implemented in our case for nonlinear classification.
Fig (3) Block Diagram of Speech Emotion Recognition System using SVM Classification
5.1 RBF (RADIAL BASIS FUNCTION) KERNEL SVM
The standard form for implementing RBF function [9] is given by the following function as shown in equation
(5):
2
) ……………….. (5)
where h(x) = hypothesis of the given data element x, Wn = weight of the given samples, ||X-Xn||2
) =
the basis function, wherein ||X- Xn||2
= the radial distance
VI. Experimental Results And Analysis
The data is trained by extracting the voice features MFCC and LPCC. The performance of speech
emotion recognition system is subjective to many factors, mainly the quality of the speech samples, the
extracted features and classification algorithm that has been used. The following figure illustrates the extracted
MFCC and LPCC features of various emotions respectively.

Fig (4) Combination of MFCC and LPCC features for Joy and Fear
Fig (5) Combination of MFCC and LPCC features for Anger and Sadness
Fig (6) Combination of MFCC and LPCC features for Boredom and Neutral
Fig (7) Combination of MFCC and LPCC features for Disgust
VII. Results For Berlin Emotional Database
The following tables illustrate the Recognition rate (%) of all the emotions (starting with only 2
emotions and then gradually increasing the number of emotions) classified by SVM.
Table 1.1: Recognition rate (%) of only 2 emotions (Anger & Sad)
Recognition (%) Linear SVM Non linear SVM
MFCC 87.83 100
LPCC 91.01 100
MFCC+LPCC 90.48 100

Table 1.2: Recognition rate (%) of only 3 emotions (Anger & Sad & Joy)
Recognition (%) Linear SVM Nonlinear SVM
MFCC 53.85 100
LPCC 53.46 100
MFCC+LPCC 61.54 100
Table 1.3: Recognition rate (%) of only 4 emotions (Anger & Sad & Joy & Fear)
Recognition (%) Linear SVM Nonlinear
SVM
MFCC 46.06 85.45
LPCC 53.33 87.88
MFCC+LPCC 56.06 86.36
Table 1.4: Recognition rate (%) of 5 emotions (Anger & Sad & Joy & Fear & Disgust)
MFCC 46.67 84.53
LPCC 45.60 85.87
MFCC+LPCC 47.20 86.13
Table 1.5: Recognition rate (%) of 5 emotions (Anger & Sad & Joy & Fear & Disgust)
MFCC 46.67 84.53
LPCC 45.60 85.87
MFCC+LPCC 47.20 86.13
Table 1.6: Recognition rate (%) of 6 emotions (Anger & Sad & Joy & Fear & Disgust & Boredom)
MFCC 42.76 83.33
LPCC 41.23 82.24
MFCC+LPCC 46.27 87.50
Table 1.7: Recognition rate (%) of 7 emotions (Anger & Sad & Joy & Fear & Disgust & Boredom & Neutral)
MFCC 35.70 84.30
LPCC 38.88 60.00
MFCC+LPCC 40.00 88.04
Table 1.8 Confusion Matrix of MFCC+LPCC feature extracted using RBF kernel SVM:
Anger Joy Sadness Fear Disgust Boredom Neutral
Anger 94.30 0 0 4.06 0 0 1.62
Joy 0 90.47 0 9.52 0 0 0
Sadness 0 0 95.00 0 0 0 5.00
Fear 19.69 0 0 75.75 1.51 3.03 0
Disgust 0 0 0 4.87 95.12 0 0
Boredom 0 0 2.46 0 0 97.53 0
Neutral 3.89 0 1.29 0 0 0 94.80
Table 1.9 Confusion Matrix of MFCC+LPCC feature extracted using Linear SVM:
Anger 79.67 0 4.06 2.43 0 9.75 4.06
Joy 68.85 0 8.19 6.55 0 14.75 1.63
Sadness 28.57 0 41.07 1.78 0 10.71 17.85
Fear 31.74 0 6.34 44.44 0 4.76 12.69
Disgust 64.10 0 15.38 10.25 5.12 0 5.12
Boredom 31.94 0 5.55 2.77 0 43.05 16.66
Neutral 32.85 0 2.85 7.14 1.42 10.00 45.71

VIII. Results For Savee Emotional Database
The SAVEE emotion speech database has been considered with 7 different emotions such as happiness,
angry, sad, surprise, fear, disgust and neutral. This database has been recorded as a necessity for the
improvement of an automatic emotion recognition system. This database comprises of recordings from 4 male
actors in 7 different emotions. The 4 different speakers are Darren Cosker (DC - 01), James Edge (JE - 02), Joe
Kilner (JK - 03) and Kevin Lithgow (KL -04). The following tables illustrate the emotion recognition rate for 7
different emotions.
Table 1.10: Emotion Recognition rate (%) of Speaker 01 – DC
MFCC 59.17 91.67
LPCC 59.17 86.67
MFCC+LPCC 73.33 92.50
Table 1.11: Emotion Recognition rate (%) of Speaker 02 - JE
MFCC 71.67 93.33
LPCC 76.67 78.33
MFCC+LPCC 50.83 88.33
Table 1.12: Emotion Recognition rate (%) of Speaker 03 - JK
MFCC 51.67 77.50
LPCC 51.67 75.83
MFCC+LPCC 62.50 81.67
Table 1.13: Emotion Recognition rate (%) of Speaker 04 - KL
MFCC 36.67 97.50
LPCC 51.67 80.83
MFCC+LPCC 35.00 84.17
The following table 1.14 illustrates the Confusion Matrix for all 7 emotions of SAVEE Database for Speaker 01
- DC using linear kernel.
Table 1.14: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 01 - DC using Linear Kernel
Emotions Anger Joy Sadness Fear Disgust Boredom Neutral
Anger 75 0 0 0 0 16.66 8.33
Joy 0 100 0 0 0 0 0
Sadness 0 0 0 0 0 11.11 88.88
Fear 6.66 0 0 85.71 0 0 14.28
Disgust 0 0 0 0 66.66 0 33.33
Boredom 0 0 0 0 6.66 93.33 0
Neutral 0 0 0 0 0 0 100
The following table 1.15 illustrates the Confusion Matrix for all 7 emotions of SAVEE Database for Speaker 01
- DC using Non Linear kernel.
Table 1.15: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 01 - DC using Non Linear
Kernel
Emotions
Anger 64.28 0 0 0 7.14 7.14 21.42
Joy 0 100 0 0 0 0 0
Sadness 0 0 100 0 0 0 0
Fear 0 0 0 100 0 0 0
Disgust 0 0 0 0 100 0 0
Boredom 0 0 0 0 0 100 0
Neutral 0 0 0 0 0 0 100

Table 1.16: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 02 - JE using Linear Kernel
Emotions
Anger 25 0 0 0 0 0 75
Joy 0 78.57 0 0 0 0 21.42
Sadness 0 0 100 0 0 0 0
Fear 0 0 0 20 10 0 70
Disgust 0 10 40 0 0 0 50
Boredom 0 0 0 0 0 11.11 88.88
Neutral 0 0 0 0 0 0 100
Table 1.17: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 02 - JE using Non Linear Kernel
Emotions
Anger 72.72 0 0 0 0 0 27.27
Joy 0 69.23 0 0 0 0 30.76
Sadness 0 0 100 0 0 0 0
Fear 0 0 0 100 0 0 0
Disgust 0 0 0 0 100 0 0
Boredom 0 0 0 0 0 100 0
Neutral 0 0 0 0 0 0 100
Table 1.18: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 03 - JK using Linear Kernel
Emotions
Anger 76.92 0 0 0 0 0 23.07
Joy 13.33 80.00 0 0 0 0 6.66
Sadness 0 0 7.14 0 0 0 92.85
Fear 0 0 0 53.84 0 7.69 38.46
Disgust 6.66 0 0 0 20 0 80
Boredom 13.33 0 0 0 0 86.66 0
Neutral 0 0 0 0 0 0 100
Table 1.19: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 03 - JK using Non Linear Kernel
Anger 86.66 0 0 0 0 0 13.33
Joy 6.66 93.33 0 0 0 0 0
Sadness 0 0 100 0 0 0 0
Fear 0 0 0 64.28 0 0 35.17
Disgust 0 0 0 14.28 85.71 0 0
Boredom 0 0 0 0 0 38.46 61.53
Neutral 0 0 0 0 0 0 100
Table 1.20: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 04 - KL using Linear Kernel
Emotions
Anger 0 0 0 0 0 0 100
Joy 0 100 0 0 0 0 0
Sadness 0 0 0 0 0 0 100
Fear 0 0 0 0 0 0 100
Disgust 0 0 0 0 14.28 0 85
Boredom 0 0 0 0 0 0 100
Neutral 0 0 0 0 0 0 100
Table 1.21: Confusion Matrix of MFCC + LPCC feature extracted for Speaker 04 - KL using Non Linear
Kernel
Anger 33.33 0 0 0 0 0 66.6
Joy 0 100 0 0 0 0 0
Sadness 0 0 100 0 0 0 0
Fear 0 0 0 100 0 0 0
Disgust 0 0 0 0 100 0 0
Boredom 0 0 0 0 0 100 0
Neutral 0 0 0 0 0 0 100

IX. Conclusion
In conclusion, experiments show that the highest accuracy rate for Speech emotion Recognition has
been achieved using RBF kernel. The table noticeably depicts that emotion is highly recognized accurately
where nonlinear SVM (RBF kernel) is used as compared to that of linear SVM. It provides 100 % accuracy
when only 2 to 3 emotions are supplied to nonlinear SVM. But as an when the number of emotions goes on
increasing, the accuracy level slowly and gradually goes down. In case of Linear SVM, even 2 emotions results
in fair accuracy and when more number of emotions are added unto it, the recognition rate yet goes down to less
than even 50 percent.
Accuracy of the developed model is presented using Confusion Matrix. The Recognition result of the
combined MFCC and LPCC extracted features are compared with the isolated results. The maximum accuracy
rate that reaches by using the combination of feature extraction method is 88.59% for nonlinear SVM and 37.38
% for linear SVM. From the above analysis, it is clear that the proposed system gives different accuracy level
for both the databases. The system works fine for the Berlin Database as compared to SAVEE database. The
only drawback of this system is that the Emotion Recognition rate of Linear SVM turns out to be too low. The
future task that can be done is to improve the accuracy rate when using a linear SVM.
References
[1]. Igor Bisio, Alessandro Delfino, Fabio Lavagetto, Mario Marchese, And Andrea Sciarrone, “Gender-Driven Emotion Recognition
Through Speech Signals For Ambient Intelligence Applications”, Ieee Transactions On Emerging Topics In Computing, Digital
Object Identifier 10.1109/Tetc.2013.2274797, 21 January 2014.
[2]. http://www.expressive-speech.net/, Berlin emotional Speech database
[3]. http://personal.ee.surrey.ac.uk/Personal/P.Jackson/SAVEE/, SAVEE Database
[4]. Nitin Thapliyal, Gargi Amoli, „Speech based Emotion Recognition with Gaussian Mixture Model‟, International Journal of
Advanced Research in Computer Engineering & Technology, Volume 1, Issue 5, July 2012, ISSN: 2278 1323.
[5]. Inma Mohino-Herranz, Roberto Gil-Pita, Sagrario Alonso-Diaz and Manuel Rosa-Zurera, “MFCC Based Enlargement Of The
Training Set For Emotion Recognition In Speech”, Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.1,
February 2014.
[6]. Yixiong Pan, Peipei Shen and Liping Shen, “Speech Emotion Recognition Using Support Vector Machine”, International Journal of
Smart Home, Vol. 6, No. 2, April, 2012.
[7]. T.-L. Pao, Y.-T. Chen, J.-H. Yeh, P.-J. Li, “Mandarin emotional speech recognition based on SVM and NN”, Proceedings of the
18th International Conference on Pattern Recognition (ICPR?06), vol. 1, pp. 1096-1100,September 2006.
[8]. Showkat Ahmad Dar, Zahid Khaki „ Emotion Recognition based on Audio Speech‟, IOSR Journal of Computer Engineering (IOSR-
JCE) ),e-ISSN: 2278-0661, p-ISSN: 2278-8727, Volume 11, Issue 6 (May - Jun 2013), PP 46-50.
[9]. Yaser Abu-Mostafa, “Radial Basis Function”, Learning from data: Introductory Machine Learning, Hameetman Auditorium,
Caltech, May 24, 2012.

A017410108

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to A017410108 (20)

More from IOSR Journals (20)

Recently uploaded (20)

A017410108