Dynamic Spectrum Derived Mfcc and Hfcc Parameters and Human Robot Speech Interaction

ACEEE Int. J. on Electrical and Power Engineering, Vol. 02, No. 02, August 2011

Dynamic Spectrum Derived Mfcc and Hfcc
Parameters and Human Robot Speech Interaction
Krishna Kumar Sharma1, Piyush Kapoor1, Prof. G C Nandi1, and Dr Pavan Chakraborty1
1
Indian Institute of Information Technology, Allahabad, India
Email: krisshna.sharma@gmail.com, piyushkapoor7@yahoo.com, gcnandi@iiita.ac.in, pavan@iiita.ac.in

Abstract—Using the M el-frequency cepstral coefficients dynamic HFCC i.e. HFCC is the spectral filtered cepstral
(MFCC), Human Factor cepstral coefficients (HFCC) and coefficient in the log spectral domain. And another feature is
their new parameters derived from log dynamic spectrum and derived from log dynamic spectrum i.e. sHFCC. Now, three
dynamic log spectrum, these features are widely used for
set of feature vector are used to test recognition performance,
speech recognition in various applications. But, speech
recognition systems based on these features do not perform
those are respectively: 1) MFCC + sMFCC + MFCC,
efficiently in the noisy conditions, mobile environment and 2) HFCC + sHFCC + HFCC, and 3) MFCC + HFCC
for speech variation between users of different genders and + MFCC + HFCC. Among these feature set number three
ages. To maximize the recognition rate of speaker independent performed best in recognition percentage with 86.17% in the
isolated word recognition system, we combine both of the above lab environment (closed window), 82.33% in lab environment
features and proposed a hybrid feature set of them. We tested (open window) and 73.67% in outdoor noisy environment.
the system for this hybrid feature vector and we gained results Feature set number seven also performed good but here we
with accuracy of 86.17% in clean condition (closed window), have to filter data in two filter HFCC filter bank and MFCC
82.33% in class room open window environment, and 73.67%
filter bank, so it take more time to process data as compare to
in outdoor with noisy environment.
other feature set. After extracting features from the speech
Index Terms— MFCC, HFCC, HMM, HOAP-2
samples, we need to generate codebook from features. Linde–
Buzo–Gray algorithm [4] is used to quantize features, this is
a iterative technique of vector quantization. And then Hidden
I. INTRODUCTION Markov Model (HMM) [5] technique is used to get good
recognition result. To test the speech recognition system in
In the present scenario, robotics are gaining an increasing the real time, a 25 degree of freedom humanoid robot HOAP-
role in the social life. Speech is a most natural way to 2 is used [6], This HOAP-2 is simulated in the WEBOTS real
communicate for human as compare to eye-gazing, facial time simulation software [7]. The rest paper is organized as
expression, and gestures to interact with robot [1]. But speech follows. In section 2, we describe speech recognition (SR)
recognition performance varies according to environment and system, which contains techniques to extract different
users. Robots are mobile in nature and controlled by different cepstral coefficients, vector quantization method, and HMM
users so, it should be noise robust, environment adaptability, model design method. In section 3, we present proposed
and user adaptability for different ages and sex. To achieve method. Section 4, describes results and comparison. And,
noise robustness, environment adaptability, and user Section 5 concludes the paper.
adaptability, features MFCC, sMFCC,and MFCC [2] are
widely used as a feature vector for speech recognition. HFCC II. SR SYSTEM
[3] is also has been used as a feature for speech recognition.
HFCC outperform in the clean condition, but to make it MFCC is widely used speech feature for automatic speech
efficient feature vector in the noisy condition we used dynamic recognition. The functionality of MFCC is attributed to
cepstral coefficient of HFCC. As described above feature characteristics of the triangular sized filter bank as shown in
vector of MFCC are used for recognition and similarly HFCC Fig. 2. Calculated energy of each filter smoothes the speech
are also separately used for recognition purpose, both spectrum, repressing the effects of pitch, and the warped
features work smartly in different-different situations. MFCC frequency scale provides changeable sensitivity to the speech
filter bank and HFCC filter bank are different in design spectrum. But MFCC does not resemble the approximate
perspective, in MFCC filter bank spacing is dissociated with critical bandwidth of human auditory system. HFCC is
filter bandwidth but, in HFCC filter spacing is associated devised to dissociate filter bandwidth from number of filters
with equivalent rectangle bandwidth (ERB) that is introduced and frequency range. When signal noise ratio is high, then
by Moore and Glasberg. Static MFCC and static HFCC MFCC and HFCC performed well, but in the noisy situation
features can attain high accuracy in the clean environment their performances degrade. To reduce the noise effect in the
but, in case of robot environment, it is not always clean. It signal, we calculate their dynamic and static parameter of
varies and has noise. So, to tune parameters with above MFCC and HFCC. MFCC and HFCC are extracted as given in
mentioned conditions, dynamic parameters are used of the Fig. 1. MFCC and HFCC differ only in the filter design
cepstral coefficients. Dynamic MFCC i.e. MFCC is the technique. We used different combination of parameters.
spectral filtered cepstral coefficient in the log spectral domain. MFCC and its dynamic features and in the similar way HFCC
And another feature is derived from log dynamic spectrum and its dynamic features are used as shown in the Fig. 1. And
i.e. sMFCC. And another updated feature is HFCC, and its combined features of MFCC and HFCC are used as a new
20
© 2011 ACEEE
DOI: 01.IJEPE.02.02.69


proposed feature for this application as shown in the Fig. 1.
A. Feature Extarction
To calculate cepstral coefficients following steps are used
as per shown in the Fig. 1. Let s(n) is the input speech signal
recorded at 16KHz frequency with signed 16- bit resolution.
1) Signal s(n) is put through a low order digital system. to
spectrally flatten the signal and to make less susceptible to
finite precision effect later in the speech processing.
, where a=0.95 (1)
2) Speech signal is quasi-static signal, so it is divided in to
frames of 25msec length for 16kHz speech signal [9].
In other words, we can also say that a frame contain FL=400
samples and next frame start followed by 160 samples or
adjacent frames are being separated by 160 samples.
3) Now, framed signals are passed through hamming
window to maintain continuity in the signal. There are other
windows also but it efficiently remove side ripples in the
signal.

, where τ is frame index, FL is the frame length.
4) Now, to get frequency spectrum of the signal, FFT is
performed on the windowed signal.

(3)
5) Hearing perception of human is not equally sensitive to all Figure. 1 Block Diagram from the feature extraction to design
frequencies signal, it behave logarithmically to high HMM model.
frequencies signal and linearly to low frequencies signal. Here, is the center frequency of the filter and
Filters are designed according to mel scale to the spectrum. is a group of triangular filters that have equal-height. Boundary
Mel scale is approximately logarithmic for above 1kHz and points are uniformly spaced in the mel scale. MFCC filter
linearly below 1kHz.mel frequency is calculated from bank is obtained as shown in the Fig. 2 for the 19 filters in the
following equation: range of 8KHz.
6) The output of the mel filtering is passed to logarithmic
function (natural logarithm) to obtain log-energy output.
is denoted in the form of K, fs, and FL as follows:
(5)
Here, FL is the frame length.
And, MFCC filter bank is designed as follows:

Figure 2. MFCC Filter bank of 19 filter in 8KHz range.

21
© 2011 ACEEE
DOI: 01.IJEPE.02.02. 69


7) To obtain the static feature of MFCC Discrete Cosine samples feature vector from LBG technique. And from these
Transform (DCT) is employed to log-energy. 36 samples we designed HMM model of each word. And
finally, performing a viterbi search algorithm to find out most
likely state sequence in HMM given a sequence of observed
output.

III. PROPOSED METHOD
j is the index of the cepstral coefficient. M is the number of
filter bank that are used and, J is the number of MFCC are After successfully extracting features from MFCC
used. This obtain MFCC are called static feature of speech, coefficient and its dynamic coefficient are used as features.
these are less immune to noise and changing environment Similarly HFCC coefficient and its dynamic coefficient are
condition. To make it more robust feature are extracted from also used as feature vector. As shown in the Fig. 1, the key
dynamic spectrum as per shown in the figure 1. difference between these two parameters is in the design of
8) HFCC features are also obtained by employing HFCC filter the filter bank that is described in the ASR system. To increase
bank in place of MFCC filter bank in the feature extraction system efficiency we proposed possible combination of the
step 6. HFCC filter is designed as per described by Mark D. features, which include both filter’s characteristics, in the
Skowronski and John G. Harris [9]. That will be static, following steps:
dynamic log spectrum parameter and log dynamic spectrum Step 1: as described in the section 2.1, we obtained MFCC
parameter by applying further steps from 6 to 7. and HFCC features and their dynamic parameters also.
9) After these steps, we got following features from the HFCC Step 2: according our proposed method, we make
filter bank and MFCC filter bank: combined feature vector of both parameter and vector
MFCC, MFCC, sMFCC, MFCC, HFCC, HFCC, quantized the parameter.
ggsHFCC, HFCC. Now, we need to generate codebook Step 3: and from generated codebook after vector
for the different feature combinations as described in the quantization, we develop HMM model of the each word.
INTRODUCTION part. codebook is generated from the . Step 4: system is trained from step 3, now to use this
Linde–Buzo–Gray (LBG) vector quantization algorithm. 12 system we find out testing samples maximum likelihood from
coefficients of each feature are used for each frame in the the HMM models using viterbi algorithm.
further processing, in the 6 type parameter set 3 features are HMM model is not similar to markov model whose states
used to make feature vector but, in one feature vector 4 can not be directly observed [10]. HMM model can be
features are used to make feature vector. described in to two sections; 1) one discusses about the
entities associated with the HMM model and, 2)present
B. Vector Quantization technique of HMM based similarity measure.
The LBG vector quantization is an iterative technique of HMM contain following entities:
quantization. The codebook is generated on the binary- 1. is finite state sequence, where
splitting method, means initially average code vector is split N is the number of states.
in to two, and further in to vector, n is the splitting number. . 2. , denoting the transition
Vector quantization is used to generate codebook for
probability from . And . (10)
HMM. Following methodology is used to quantize vector:
1. Initially, calculate the centroid of each frame for the speech 3. Emission matrix , denoting the probability of
sample. emission symbol O when state is S.
2. Now, split the centroid of codebook Yn according to: 4. Initial state probability distribution

Where (9)
3. Find out the nearest – neighbor for each training vector. denoting the probability of state
This is done using the K-mean iterative algorithm. HMM based similarity measure:
4. Update the centroid according to member of the cluster. HMM is trained for each utterance by baum-Welch
5. Repeat step 3 and 4 until the average distance falls below algorithm [5]. HMM find out similarity between two
a preset threshold. utterances by calculating similarity of
6. Repeat steps 2, 3, and 4 until codebook of size M is reached.
as:
M= , n is the splitting number, and M is the desired size of
the codebook.
C. Hidden Markov Model (Hmm) Where, is calculated using the famous viterbi
After vector quantization of samples we get codebook to algorithm, in which log likelihood is used.
generate HMM model of words. Each word is spoken by 18
people in 10 different conditions and manner, so each word
has 180 samples. Now, each word’s 180 samples are vector
quantized by LBG technique. From 180 samples we get 36
22
© 2011 ACEEE
DOI: 01.IJEPE.02.02.69


IV. EXPERIMENT AND RESULTS Form this experiments we get good results for our own dataset
that is described in the following tables Table 1, Table 2, and
The design of the Speech recognition system is varied
Table 3. In the Table 1, we described about result obtained
according to their objective, like isolated speech recognition
from the MFCC, sMFCC, and MFCC features. In the
system, and continuous speech recognition system for two
Table 2, we described about result obtained from the
modes speaker dependent and speaker independent.
HFCC, sHFCC, and HFCC features. And In the Table 3,
TABLE I. we described about result obtained from the MFCC,
FEATURE VECTOR 1 HFCC, MFCC, and HFCC features. From the obtained
result we found that in the clean environment of the lab in the
closed window form then recognition rate is less varied for
all three feature vector, that are respectively: 85.33%, 85.67%,
and 86.17% for Table 1, Table 2 and Table 3. And if we in-
crease the noise ratio in the signal than this recognition rate
change slightly. Now, we evaluated in the class room noisy
or open window form, obtained results are as followed 80.33%,
81.33%, and 82.33% from three tables respectively Table 1,
Table 2, and Table 3. And we also tested for the real time
environment in the outdoor noisy environment for the all
feature sets, obtained results are as followed 69.33%, 70.17%,
and 73.67% from Table 1, Table 2, and Table 3. From ob-
tained results, we can find out that feature set of combined
filter work better as compare to the single filtered feature
sets. This is like features are facilitated with the characteris-
tics of the both filters and they efficiently resemble human
auditory system because MFCC filter bank works linearly for
below 1 KHz signal and logarithmic for above 1 KHz signal.
For the training of the isolated speech recognition system of And HFCC filter bank also resemble human auditory system
the speaker independent mode, we gathered speech samples with modified filter bank spacing than MFCC.
of the different age people, different environment (lab room
with close window, lab room with open window or class room TABLE III.
FEATURE VECTOR 2
environment, and outdoor noisy environment) [11]. To test
this recognition system, we used humanoid robot HOAP-2.
In the real time we commanded humanoid robot and observed
results. We collected each word spoken by 18 different users
10 times by each user or in other words 180 times each word.
We categorized users in to the three age categories 6-13 years,
14-50 years, and 51-70 years of both genders. We processed
the collected data and trained the system and tested it for the
other speaker other than the trained sample’s speakers.
TABLE II.
FEATURE VECTOR 3

V. CONCLUSION AND FUTURE WORK
In This paper, we described about different features based
speaker independent isolated speech recognition system, and
find out there efficiency in the different mobile environment
and user adaptability. Features extraction only differ in the
filtering process based on their filter bank construction, and
we find out that efficiency of recognition increase as we
combine different features and their differentiated parameters
23
© 2011 ACEEE
DOI: 01.IJEPE.02.02. 69


as compare to one filtered parameter. Combined features set [ 4 ] WE B , ” h t t p : / / w w w. s o f t w a r e p r ac t i ce . o rg / w ik i /
of MFCC and HFCC take more computing time in the training Vector_Quantization_LBG_Pseudo_Code”
phase and testing as compare to individual features set of [5] L. Rabiner. A tutorial on Hidden Markov Models and selected
applications in speech recognition.Proc. of IEEE, 77(2):257-
MFCC and also from HFCC features set. But, results improved
286,1989.
in the combined feature set as shown in the Table 1, Table 2, [6] WEB, “http://robita.iiita.ac.in/hoap2instruction03e.pdf”
and Table 3. In the future work, we would like to use this [7] WEB, “http://www.cyberbotics.com/”
technique for large number of words, the continuous speech [8] Iosif Mporas, Todor Ganchev, Mihalis Siafarikas, Nikos
recognition, and would like to make it more robust for the Fakotakis, “Comparison of Speech Features on the Speech
different races people also. Recognition Task,” Journal of Computer Science 3 (8): 608-616,
2007.
REFERENCES [9] ETSI ES 201 108, V1.1.3 (2003-09). ETSI standard: speech
processing, transmission and quality aspects (STQ); Distributed
[1] Carlos Toshinori Ishi, Shigeki Matsuda, Takayuki Kanda, speech recognition; Front-end feature extraction algorithm;
Takatoshi Jitsuhiro, Hiroshi Ishiguro, Satoshi Nakamura, and Compression algorithms; (Sect. 4, pp. 8–12).
Norihiro Hagita, “A Robust Speech Recognition System for [10] M.Gales and SJ Young.The Application of Hidden Markov
Communication Robots in Noisy Environments,” IEEE Models in Speech Recognition. Foundations and Trends in Signal
Transactions on robotics, vol. 24, no. 3, June 2008. Processing- 1(3), pp. 195-304,2008.
[2] Nengheng Zheng, Xia Li, Houwei Cao, Tan Lee, and P. C. [11] Hong Liu, Xiaofei Li, “A Selection Method of Speech
Ching,”Deriving MFCC parameters from the dynamic spectrum Vocabulary for Human-Robot Speech Interaction” Systems Man
for robust speech recognition”, Chinese Spoken Language and Cybernetics (SMC), 2010 IEEE International Conference on ,
Processing, 2008. ISCSLP ’08. 6th International Symposium on. vol., no., pp.2243-2248, 10-13 Oct. 2010
[3] Mark D. Skowronski and John G. Harris, “Improving the filter
bank of a classic speech feature extraction algorithm,” IEEE Intl
Symposium on Circuits and Systems, Bangkok, Thailand, vol IV,
pp 281-284, May 25 - 28, 2003.

24
© 2011 ACEEE
DOI: 01.IJEPE.02.02. 69

Dynamic Spectrum Derived Mfcc and Hfcc Parameters and Human Robot Speech Interaction

More Related Content

What's hot (20)

Similar to Dynamic Spectrum Derived Mfcc and Hfcc Parameters and Human Robot Speech Interaction (20)

More from IDES Editor (20)

Recently uploaded (20)

Dynamic Spectrum Derived Mfcc and Hfcc Parameters and Human Robot Speech Interaction