INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 4, ISSUE 06, JUNE 2015

ISSN 2277-8616

Speech-To-Text Conversion (STT) System Using

Hidden Markov Model (HMM)
Su Myat Mon, Hla Myo Tun
Abstract: Speech is an easiest way to communicate with each other. Speech processing is widely used in many applications like security devices,
household appliances, cellular phones, ATM machines and computers. The human computer interface has been developed to communicate or interact
conveniently for one who is suffering from some kind of disabilities. Speech-to-Text Conversion (STT) systems have a lot of benefits for the deaf or
dumb people and find their applications in our daily lives. In the same way, the aim of the system is to convert the input speech signals into the text
output for the deaf or dumb students in the educational fields. This paper presents an approach to extract features by using Mel Frequency Cepstral
Coefficients (MFCC) from the speech signals of isolated spoken words. And, Hidden Markov Model (HMM) method is applied to train and test the audio
files to get the recognized spoken word. The speech database is created by using MATLAB.Then, the original speech signals are preprocessed and
these speech samples are extracted to the feature vectors which are used as the observation sequences of the Hidden Markov Model (HMM)
recognizer. The feature vectors are analyzed in the HMM depending on the number of states.
Keywords: Speech Recognition, End Point Detection, MFCC, HMM, MATLAB

1. Introduction
Human interact with each other in several ways such as facial
expression, eye contact, gesture, mainly speech. The speech
is primary mode of communication among human being and
also the most natural and efficient form of exchanging
information among human in speech [1]. Speech-to-text
conversion (STT) system is widely used in many application
areas. In the educational field, STT or speech recognition
system is the most effective on deaf or dumb students. The
recognition of speech is one the most challenges in speech
processing. Speech Recognition can be defined as the
process of converting speech signal to a sequence of words
by means of Algorithm implemented as a computer program
[1]. Basically, speech to text conversion (STT) system is
distinguished into two types, such as speaker dependent and
speaker independent systems [2]. This paper presents the
speaker dependent speech recognition system. Speech
recognition is very complexity case when processing on
randomly varying analogue signal such as speech signals.
Thus, in speech recognition system, feature extraction is the
main part of the system. There are various methods of feature
extractions. In recent researches, many feature extraction
techniques are commonly used such as Principal Component
Analysis (PCA), Linear Discriminant Analysis (LDA),
Independent Component Analysis (ICA), Linear Predictive
Coding (LPC), Cepstral Analysis and Mel-frequency cepstral
(MFCCs), Kernal based feature extraction based approach,
Wavelet Transform and spectral subtraction [3]. In this paper,
Mel Frequency Cepstral Coefficients (MFCC) method is used.
It is based on the characteristics of the human ear's hearing,
which uses a nonlinear frequency unit to simulate the human
auditory system. Mel frequency scale is widely used to extract
features of the speech. Mel-frequency cepstral features
provide the rate of recognition to be efficient for speech
recognition as well as emotion recognition system through
speech [4]. Moreover, Vector Quantization (VQ), Artificial
Neural Network (ANN), Hidden Markov Model (HMM),
Dynamic Time Warping (DTW) and various techniques are
used by the researchers in recognition. Among them, HMM
recognizer is currently dominant in many applications.
Nowadays, STT system is fluently used in many control
systems, mobile phones, computers and so forth. Therefore,

speech recognition system is more and more popular and

useful in our daily lives. In the system, MFCC and HMM are
implemented by MATLAB.

2. Methodology
A. End Point Detection
Classification of speech into voiced or unvoiced sounds
provides a useful basis for subsequent processing. A threeway classification into silence/unvoiced/voiced extends the
possible range of further processing to tasks such as stop
consonant identification and endpoint detection for isolated
utterances [5]. In noisy environment, speech samples
containing unwanted signals and background noise are
removed by end point detection method. End point detection
method is based on the short-term log energy and short-term
zero crossing rate [6]. The logarithmic short-term energy and
zero crossing rates are calculated in the following equation [1]
and [2]
2
N
n=1 log(s(n)

Elog =
ZCR=

1
2

N
n=1 |sgn

sgn s n =

+1
-1

[1]

s n+1 -sgn s n |

[2]

s n 0
s(n)<0

Wheres(n) is the speech signal, Elog is the logarithmic shortterm energy and ZCR is the short-term zero crossing rate.
B. Mel Frequency Cepstral Coefficient (MFCC)
Feature extraction is the most important part of the entire
system. The aim of feature extraction is to reduce the data
size of the speech signal before pattern classification or
recognition. The steps of Mel frequency Cepstral Coefficients
(MFCCs) calculation are framing, windowing, Discrete
Fourier Transform (DFT), Mel frequency filtering, logarithmic
function and Discrete Cosine Transform (DCT).Fig.1 shows
the block diagram of MFCC process.

349
IJSTR2015
www.ijstr.org

INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 4, ISSUE 06, JUNE 2015

probability (A) and symbol emission probability (B). In HMMbased speech recognition system, there exist three main
problems called evaluation, decoding and learning problems.
The training and testing algorithm of HMM are discussed in
details [8]. The probability of observations or likelihood given
the model determines the expected recognized word. It is
calculated by the following equation [3]

Speech

Framing

Windowing

ISSN 2277-8616

Discrete Fourier
Transform (DFT)

L=P O =

N
i=1 T (i)

[3]

MFCC
Discrete Cosine
Transform (DCT)

Logarithmic

Where P O is the probability of observations addressed by

forward algorithm. N is the number of states and T (i) is the
forward variables with the length of observations. The highest
probability of observations determines the recognized spoken
word.

Mel Frequency
Filtering

Fig .1. Block diagram of MFCC

Framing: It is the first step of the MFCC. It is the process of
blocking of the speech samples obtained from the analogue to
digital conversion (ADC) of the spoken word, into the number
of frame signal with 20- 40ms frame time length. Overlapping
is needed to avoid loss of information. Windowing: In order to
reduce the discontinuities at the start and end of the frame or
to be smooth of the first and last points in the frame,
windowing function is used. DFT: Discrete Fourier Transform
(DFT) is used as the Fast Fourier Transform (FFT) algorithm.
FFT converts each frame of N samples from the time domain
into the frequency domain. The calculation is more precise in
frequency domain rather than in time domain. Mel frequency
filtering: The voice signal does not follow the linear scale and
the frequency range in FFT is so wide. It is perceptual scale
that helps to simulate the way human ear works. It
corresponds to better resolution at low frequencies and less at
high. Logarithmic function: Logarithmic transformation is
applied to the absolute magnitude of the coefficients obtained
after Mel-scale conversion. The absolute magnitude operation
discards the phase information, making feature extraction less
sensitive to speaker dependent variations. DCT: Discrete
cosine transform (DCT) converts the Mel-filtered spectrum
back into the time domain since the Mel Frequency Cepstral
Coefficients are used as the time index in recognition stage.

3. Implementation
The flowchart of speech to text conversion is illustrated in Fig
.2. To convert input speech to text output, the four main steps
are developed by using MATLAB.These steps are speech
database, preprocessing, feature extraction and recognition.
Firstly, five audio files are recorded with the help of computer.
Each audio file contains ten different pronunciation audio files.
So, there are total of fifty audio files are recorded in speech
database. The speech signals at low frequencies have more
energy than at high frequencies. Therefore, the energies of
signal are necessary to be boost at high frequencies.
According to the saturation of environment, the unwanted
noise may affect the recognition rate worse. This problem can
be overcome by end point detection method. After
preprocessing stage is finished, the speech samples are
extracted to features or coefficients by the use of Mel
Frequency Cepstral Coefficient (MFCC). Finally, these MFCC
coefficients are used as the input of Hidden Markov Model
(HMM) recognizer to classify the desired spoken word. The
desired text output can be generated by HMM method even if
the test audio file is included in the existing speech database.

C. Hidden Markov Model Recognizer

In recognition or classification of the speech signal, there are
many approaches to recognize the test audio file. The
methodologies of speech recognition are: ANN, GMM, DTW,
HMM, Fuzzy logic and various other methods. Among them,
HMM techniques are widely used in many applications than
any other ones. There are four types of HMM model used in
speech processing. Details of HMM models are given in [7].
The phonemes in speech follow the left to rightsequences, so
the structure of HMM is a left-to-right structure. The states of
HMM model represent the word or acoustic phonemes in
speech recognition. The number of states of HMM model is
randomly chosen to model. The choice of the number of states
causes to change the feature vectors or observations. It
affects the recognition rate or accuracy of speech
recognition.The most flexible and successful approach to
speech recognition so far has been Hidden Markov Models
(HMMs). HMM is the popular statistical tool for modeling a
wide range of time series data. In speech recognition area,
HMM has been applied with great success to problem such as
part of speech classification [1]. HMM word model is
composed of initial state probability (), state transition

Start

Record Speech

Removal Noise

Calculate MFCCs

Match with speech

database?
(HMM)

No

Yes
Desired Text Output

End

Fig .2. Flowchart of speech to text conversion

350

IJSTR2015
www.ijstr.org

INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 4, ISSUE 06, JUNE 2015

4. Simulation Results

ISSN 2277-8616

Original Signal (Flower)

0.3

In this HMM-based speech to text conversion system, five

audio files such as apple, banana, computer, flower and key
are modeled in HMM. The original signal at the sampling rate
of 8 kHz are demonstrated in Fig .3.

0.2

0.1
Original Signal (Apple)
0.3

Amplitude

0.2

-0.1

Amplitude

0.1

-0.2

-0.1

-0.3

-0.2

-0.4

4
5
6
Number of samples

-0.3

-0.4

10
4

x 10

Original Signal (Key)

0

3
Number of samples

0.3

x 10

0.25
Original Signal (Banana)

0.2

0.2

0.15

0.15

0.05

Amplitude

0.1

Amplitude

0.1

0.05
0

-0.05
-0.05

-0.1
-0.1

-0.15
-0.15

-0.2
-0.2

0.5

1.5

2
2.5
3
Number of samples

3.5

4.5

6
Number of samples

10

12
4

x 10

x 10

Original Signal (Computer)

Fig .3 Amplitude versus number of samples of five original

signals

0.2
0.15

In recognition, the more the number of states in HMM, the

better the recognition rate or accuracy. Tables 1, 2 and 3 show
the percentages of recognition rate for speech to text
conversion.

0.1

Amplitude

0.05
0

Table .1 Table of Percentage Accuracy for three states of

HMM (N=3)

-0.05
-0.1
-0.15
-0.2

3
Number of samples

Train data

Number of
test

Number of
correct test

Error

Percentage
of Accuracy

Apple

50

31

19

62%

Banana

50

32

18

64%

Computer

50

33

17

66%

Flower

50

31

19

62%

Key

50

30

20

60%

x 10

351
IJSTR2015
www.ijstr.org

INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 4, ISSUE 06, JUNE 2015

Table .2 Table of Percentage Accuracy for four states of HMM

(N=4)
Train data

Number of
test

Number of
correct
test

Error

Percentage
of
Accuracy

Apple

50

37

13

74%

Banana

50

36

14

72%

Computer

50

36

14

72%

Flower

50

39

11

78%

Key

50

35

15

70%

ISSN 2277-8616

The average percentage accuracy or recognition rate for the

system is illustrated in Fig .4.At the number of state (N=5), the
average accuracy is about 87.6% as the most. It is better
recognition rate than state three and four of HMM.

5. Conclusion

Train data

Number of
test

Number of
correct
test

Error

Percentage
of
Accuracy

Apple

50

43

86%

This Speech- to-Text conversion system is implemented by

using the MFCC for feature extraction and HMM as the
recognizers. In speech database, fifty audio files are recoded
and these are analyzed to get feature vectors. These features
are initially modeling in the HMM. After that, the test spoken
word is addressed by forward algorithm of HMM. From the
simulation results, it can be clearly seen that the average
recognition rate of 87.6% achieved by the number of states
(N=5) is better accuracy than any other states. But, if the
number of states is too large, there are no enough
observations per state to train the model. So, this may
degrade the performance of the system. Thus, the choice of
the number of states in the HMM also plays an important case
in recognition. In this work, the performance of the system is
more accurate and reliable by using end point detection
algorithm in preprocessing stage.

Banana

50

42

84%

Acknowledgement

Computer

50

45

90%

Flower

50

46

92%

Key

50

43

86%

Table .3 Table of Percentage Accuracy for five states of HMM

(N=5)

The author would like to thank to Dr. Hla Myo Tun, Associate
Professor and Head of the Department of Electronic
Engineering, Mandalay Technological University for his help
and for his guidance, support and encouragement.

References
In Table .1, the percentage of recognition rate for apple and
flower is 62 %. For banana, the recognition rate is slightly
increased to 64% and the recognition rate of computer have
the best result of 66%.Whereas, the accuracy of key has the
least of 60%.For number of states (N=4),the percentage of
recognition rate is increased around 70 for all audio files. This
is shown in Table .2. According to the Table .3, the number of
states (N=5) gives the better accuracy than any other states.
The recognition rate of individual spoken word is nearly from
84 to 92%.

AVERAGE PERCENTAGE ACCURACY

FOR SPEECH-TO-TEXT CONVERSION
100.00%
90.00%
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%

87.60%
73.20%
62.80%

[1] Santosh K.Gaikwad, A review on speech recognition

techniques,
International
Journal
of
Computer
Applications, Volume 10 No.3, November 2010
[2] NishantAllawadi,Speech-to-Text System for Phonebook
Automation, Computer Science And Engineering
Department Thapar University,June 2012.
[3] Sanjivani S.Bhabad, An overview of technical progress in
speech recognition, International Journal of advanced
research in computer science and software Engineering,
Volume 3, Issuse 3, March 2013
[4] Akshay S. Utane, Emotion Recognition Through Speech
Using Gaussian Mixture Model And Hidden Markov
Model, International Journal of Advanced Research in
Computer Science and Software Engineering, Volume 3,
Issue 4, April 2013
[5] Nitin N Lokhande, Voice Activity Detection Algorithm for
Speech
Recognition
Applications,
International
Conference in Computational Intelligence (ICCIA), 2011
[6] S.A.R. Al-haddad, Automatic Segmentation and Labeling
for Continuous Number Recognition, August 21-23, 2006
(pp221-224)

N=3

N=4

N=5

Average percentage of accuracy

Fig .4 Average Recognition Rates of Speech-to-Text

Conversion System

[7] D.B. Paul, Speech Recognition Using Hidden Markov

Models, The Lincoln Laboratory Journal, Volume 3,
Number 1 (l990).
[8] Mathew Magimai Doss,Using Auxiliary Sources Of
Knowledge For Automatic Speech Recognition, Computer
Science and Engineering ,2005
352
IJSTR2015
www.ijstr.org