Mel Frequency Cepstral Coefficient: A Review

Shalbbya Ali1, Dr. Safdar Tanweer2, Syed Sibtain Khalid3, Dr. Naseem Rao4

{shalbbya.ali@gmail.com1, safdartanweer@gmail.com2, s.sibtainkhalid@jamiahamdard.ac.in3 ,

naseemjmi0786@gmail.com4}

Department Of CSE, Jamia Hamdard

Abstract. Speech is the easiest and widely used mode of communication between human
beings. The development of a human computer interface for establishing a similar dialogue
between the machine and humans was the inspiration behind speech recognition systems. This
involves training a system so that voice recognition can be carried out. The machine has to be
trained using any of the speech recognition algorithms which would extract the features of the
voice and perform speech recognition. One such algorithm is the Mel-frequency Cepstral
coefficient. The paper describes all the stages of the MFCC technique along with brief
description of each process. It also gives a comparison between linear predictive coding and
MFCC technique.

Keywords: Speech Recognition System, MFCC, Mel Frequency, Feature extraction.

1 Introduction

The interaction that takes place between a human and a computer is a common form of
communication in today’s scenario. This kind of interaction can only be carried out with the help of
some hardware devices like keyboard, touchscreen, mouse etc. but humans prefer a more natural
form of communication rather than using hardware devices for the communication. Speech is the
widest and profound means of communication used by human beings. The inspiration behind
Speech Recognition Systems is the human to human interaction, voice being the biggest feature.
Speech Recognition can enable a dictation facility by which devices can be controlled and
documents can also be created. The creation of documents by speech saves time and effort made by
a human. The business organizations have been greatly benefitted by this domain. There exists a
voice technology where the callers can input their data and get a response later without actually
interacting with a live agent. It has led to cost reduction and optimization in the business domain
eliminating the cost that would be used in keeping a live agent. Non-verbal communication between
human and a computer already existed but using voice for the same has led to the development of
the area of Speech Recognition. A voice sample contains lot of information starting from gender

ICIDSSD 2020, February 27-28, New Delhi, India

Copyright © 2021 EAI
DOI 10.4108/eai.27-2-2020.2303173
and age of a person sometimes even reflecting the state of mind of the speaker. The motive of Voice
Recognition is to use some of this information and identify the speaker on one or more of these
parameters. There are many techniques that can be used to extract features from a voice sample
which may fall under spectral or cepstral domain.
Voice Recognition is a technique that identifies a voice sample from some unique properties that
maybe acoustic or phonetic. These characteristics can be represented in the form of a speech signal
see Fig. 1.

Fig. 1. The Sound wave of Speech.

The block diagram of Speaker Identification System is shown in the figure below see Fig.2. A
sample of speech is taken and after some initial processing, properties of the speech sample are
extracted. These properties maybe acoustic properties or phonetic properties. The features are
extracted from the voice sample using some feature extraction technique. There are many extraction
techniques that exist. Some of them are MFCC, LPC RASTA etc. The selection of a technique
totally depends on what feature we want to exploit, the size of sample, the size of vocabulary,
whether the speech is an isolated word or a bunch of continuous words. After the features have been
extracted, the training is being carried out on the algorithm.[1]
The data that is offered during training period is very closely related to the data that would be given
during the testing period. The data is then stored in the database for future use. During testing period,
some data closely based on the data offered during training period is given to the system. Some
processing is done on this data and pattern matching is done. If the data matches with the one stored
in the database, the system has clearly been trained else retraining would be done or the result would
show that the pattern has not been matched.[1]
 Fig. 2 Block Diagram of Speech Recognition System.

The voice recognition is divided into two categories:

1.1 Speaker Identification
1.2 Speaker Verification
1.1 Speaker Identification: It is the process by which a voice is identified from a group of speakers.
The voice which best matches with the voice which is already stored in the database is identified
and incase the voice does not match with the one stored in the storage then the new voice becomes
another voice in the database which would be matched when a new voice has to be recognized as
shown in Fig 3. [1]

Fig. 3 Block Diagram of Speaker Identification System.

1.2 Speaker Verification: It is the process which authenticates a speaker based on some property
present in the voice sample. The identity claim maybe accepted or rejected. This type of recognition
is widely used in military,aircraft and other authentication areas. The biometric identification also
falls under the same category. The voice that is authenticated is the ones whose parameters match
the most with the one present in the database as shown in Fig 4. [1]

Fig. 4 Block Diagram of Speaker Verification System.

There are many feature extraction techniques which help in extracting some features from a speech
sample .One such technique is the Mel-Frequency Cepstral Coefficient. It is the most widely used
technique for speaker recognition because of its ease of implementation and flexibility[2].

2 Feature extraction steps in MFCC

The step by step computation of MFCC is described by the block diagram in Fig 5.

2.1 Pre-Emphasis- The speech signal x(n) has to be sent through high-pass filter:

𝑦𝑦(𝑛𝑛) = 𝑥𝑥(𝑛𝑛) − 𝑎𝑎 ∗ 𝑥𝑥(𝑛𝑛 − 1) . (1)

where y(n) is the output signal and the value of a is usually between 0.9 and 1.0.The processing of
signal is done by passing it through a filter which would emphasize its higher frequency. The energy
of the signal is increased at higher frequency[2].
2.2 Frame Blocking- The speech signal is divided into small segments of 20-30ms which are
known as frames. There may be N no. of frames where each frame that is adjacent to the other frame
is segregated by M(M<N). The values of M and N typically used are 100 and 256 respectively [2].

2.3 Windowing - It is a technique usually used in Signal processing where a speech signal is
segmented into temporal fragments. The borders of the signal that are repeating themselves are
unrelated to the real world signal. Windowing function is a smooth function which goes to zero in
the extremes. The idea is to hide the discontinuity at the borders. After the application of the window
function, there is a change in the signal but the effect on signal statistics in minimized.The window
function usually used in MFCC is the Hamming Window function. It maintains the continuity of
the first and last points in the frame.

The hamming window function is defined the equation below where W (n) is the window function
[2].

𝑌𝑌(𝑛𝑛) = 𝑋𝑋(𝑛𝑛) ∗ 𝑊𝑊(𝑛𝑛) . (2)

Fig. 5 MFCC Block Diagram.

2.4 Fast Fourier Transform- It is a signal analysis technique which is used to extract and compress
some features of the speech signal without losing any relevant information so that speech processing
becomes easier. It represents the given signal in a frequency domain. All the relevant information
from the original signal is retained only the representation of the signal is changed. The time-domain
signal is converted into frequency domain signal[3].
Triangular Band-Pass Filters- The magnitude frequency response is multiplied by a set of triangular
band pass filters so that a smooth magnitude spectrum is attained.

2.5 Discrete Cosine Transform- The formula for Discrete Cosine Transform is
𝜋𝜋
𝐶𝐶(𝑛𝑛) = ∑ 𝐸𝐸𝐸𝐸 ∗ cos (𝑛𝑛 ∗ (𝑘𝑘 − 0.5) ∗ )) (3)
40

where n = 0,1,..to N and N is the number of triangular band pass filters, L is the number of mel-
scale cepstral coefficients. The DCT is carried out on the output of the band pass filters to generate
mel-scale coefficients. The frequency domain signal is transformed into time-domain signal and the
features are also termed as the mel-scale cepstral coefficients or mel-frequency cepstral coefficients
which is used for speech recognition[3].

3 Linear Predictive Coding

This technique is also widely used for voice recognition. It uses the preceding samples and
forms a linear combination. The primary aim of LPC is frame-based analysis of the input voice
signal to create observational parameters. To apply LPC and generate the functionality, input voice
signals need to be surpassed using a pre-emphasizer. Pre-emphasizer output acts as the blocking
input to frame in which the signal is distorted into N test pieces. Further, windowing is performed
in which each frame is windowed in a way that the signal interruption at the beginning and end of
each frame is reduced. Then, each frame which is windowed is self-correlated and the total auto-
correlation value is the value of the LPC calculation[4][6].For the comparison between MFCC and
LPC see Table 3.1.
 3.1 Comparison between LPC and MFCC

1. Linear prediction coefficients (LPC) MFCC is a reproduction of the human

mimic human vocal tract and offer a hearing mechanism designed to selectively
robust speech characteristic. apply the operating theory of the ear
assuming that the human voice is a good
speaker recognizer.

2. It supports single speaker and single It supports multiple speakers and multiple
language and is quick in implementation. languages and is a little slow in
implementation.
3. Small vocabulary size makes it reliable. It is reliable for vocabulary of reasonable to
high size.

4. Linear prediction technique is used to MFCC are coefficients generated from

achieve the filter parameters comparable human auditory experience on a distorted
to the vocal tract by lowering the mean frequency scale based.
square error between the input and the
projected speech.

5. It is used in speaker recognition systems These are commonly used to recognize

where the main purpose is to retrieve the speakers and understand speech.
property of the vocal tract.

6. This offers very precise estimates of Certain pattern recognition problems

speech parameters and is relatively good relating to human voice are said to be
for calculation. accurate in cepstral coefficients.

7. LPC estimates are highly sensitive to The MFCC features are not totally accurate
quantizing noise, and may not be when background noise is present and might
appropriate for generalization. not be well adapted for generalization.

8. The LPC approach is commonly used in For security purposes, MFCC is used to
music and technical companies to build classify airline reservations, numbers
mobile robots, in telephone companies, spoken in a telephone and speech
violin tonal evaluation and other recognition software.
instrumental string gadgets.
4 Literature Review

Nisha in the year 2017 in her paper discussed Speaker identification and Speaker Verification
techniques. The identification of voice falls under the category of Speaker Identification where the
given voice is compared with a set of already existing similar sample voices present in the database
and then a match is done. However, Speaker Verification is the process of authenticating a speaker
based on some parameter or in other words the speaker is identified. Text dependent and Text
independent voice recognition systems have also been discussed. Along with this, various feature
recognition techniques like Linear Predictive Coding (LPC), Perceptual Linear Prediction(PLP),
Relative Spectral Filtering(RASTA) and Mel-Frequency Cepstral Coefficient (MFCC) along with
their merits and demerits have been discussed.The merits and demerits can easily give a trade-off to
choose the feature extraction technique best suited to the current situation and time. Some
classification techniques like Hidden Markov Model (HMM), Neural Network Modelling
Approach(NN), Dynamic Time Warping (DTW) and Vector Quantization(VQ) have also been
discussed.Ayushi T. N. P et. al. (2018) have highlighted the use of technology in the field of
communication. Applications have been created that can take part in this communication. The
functions used in establishing such a link include acoustic and articulatory speech recognition,
converting speech signal to text and then again to a signal, language translation etc. The techniques
for achieving the able mentioned functionalities have also been discussed in the paper. The speech
recognition systems can be classified into various parameters like speaker, vocal sound, vocabulary
etc. The general prerequisite for any feature extraction technique involves some preprocessing
followed by feature extraction by any of the methods be it LPC,MFCC or Dynamic Time Warping.
It also brings forth the acoustic models and a basic idea about the phonemes.The pattern
classification approach includes template based, knowledge based, statistical and neural network
based approach. The speech to text conversion methods include Hidden Markov Model HMM,
Artificial Neural Network Classifier based on Cuckoo Search Optimization. The text to speech
conversion phases include Text processing, articulator synthesis, formant synthesis and
concatenative synthesis. The language translation techniques include rule based machine translation,
statistical machine translation, example based machine translation and hybrid machine translation.
The findings and issues of each functionality has been clearly mentioned and explained. In speech
to text HMM has been classified as the best technique despite of the drawback that are present in it
and in text to speech formant synthesis has been concluded as the best approach for conversion.
Hybrid translation finds itself to be popular because of its fast learning ability and data
acquisition.Ayushi Y. K. et. al. (2017) have explained various Speech Recognition Algorithms
along with the techniques and challenges involved in the same. The issue with recognition involves
the utterance approach because no two people can speak in the same way, there is always a
difference it maybe because of their age, gender, geographical location or any other reason. Some
styles of speaking are very quick and some of them are really slow. The same word pronounced in
the American accent might sound different in the British accent. All these issues have to be analyzed
before developing an algorithm. There are various types of speaker models, which may either be
text dependent or text independent. The size of vocabulary also has an impact on speech recognition
along with the background in which the word or sentence is uttered. Different speech recognition
techniques like MFCC, PLC etc have also been briefly discussed. The approaches to classification
include acoustic phonetic approach, pattern recognition approach and artificial intelligence
approach.Pattern matching approaches for speech recognition have also been discussed. The future
scope regarding large vocabulary systems and speaker independent continuous systems have also
been discussed in the paper.Shaik Riyaz, B. et. al.(2019) in their paper have used speaker recognition
for Urdu language. The dataset used in doing so consists of 250 different Urdu words which are
spoken by 20 different speakers which comprises of 8 male speakers and 12 female speakers.The
reason for using HMM and MFCC for accomplishing the same has been found to have the highest
accuracy as compared to other feature extraction techniques and models. The dataset contains varied
speakers whose accent in Urdu language varies from Kannada accent to Uttar Pradesh accent. Vector
Quantization is used to compress the size of the feature vector and improve accuracy. For
classification, Hidden Markov Model has been used. The performance is based on the accuracy of
the Urdu words spoken with turns out to be 96.4% even after using minimum no of features as
mentioned in the paper[7][8].

5 Conclusion

The Mel-Frequency Cepstral Coefficient algorithm is a good choice for speech recognition
where multiple speakers exist. The algorithm supports multiple languages, multiple speakers and a
high vocabulary size.
Moreover, the implementation of this algorithm is quite easy as it successfully extracts features from
the voice sample. In computational context MFCC is far more expensive because of the use of Fast
Fourier Transform when measuring its range. However, the noise present in the background
environment does affect the recognition and impacts the performance of the algorithm. The change
from a noisy environment to a noise free location can help in improving the results and performance.

References
[1] Nisha. Voice Recognition Technique: A Review. International Journal for Research in
Applied Science Engineering Technology (IJRASET). May 2017;Volume 5( Issue 5)
[2] Koustav Chakraborthy, Asmita Talele & Prof. Savitha Upadhya.Voice Recognition Using
MFCC Algorithm. International Journal of Innovative Research in Advanced Engineering
(IJIRAE). November 2014;Volume 1 (Issue 10).
[3] Parvinder Pal Singh,Pushpa Rani.An Approach to Extract Feature Using MFCC.IOSR
Journal of Engineering (IOSRJEN). August. 2014;Volume 04, (Issue 08).
[4] Ayushi Y.Vadwala, Krina A. Suthar, Yesha A. Karmakar, Nirali Pandya.Survey Paper on
Different Speech Recognition Algorithm:Challenges and Techniques. International
Journal of Computer Applications.October 2017 Volume 175.
[5] Ayushi Trivedi, Navya Pant, Pinal Shah, Simran Sonik, Supriya Agrawal. Speech to text
and text to Speech Recognition Systems-A Review.IOSR Journal of Computer
Engineering(IOSR-JCE). March-April 2018;Volume 20(Issue 2).
[6] Ayushi Y.Vadwala, Krina A. Suthar, Yesha A. Karmakar, Nirali Pandya.Survey Paper on
Different Speech Recognition Algorithm:Challenges and Techniques. International
Journal of Computer Applications.October 2017 Volume 175.
[7] Riyaz, Shaik et al. Automatic Speaker Recognition System in Urdu using MFCC &
HMM.International Journal of Recent Technology and Engineering (IJRTE) February
2019; Volume-7(Issue-5S4).
[8] F. Leu and G. Lin.An MFCC-Based Speaker Identification System.IEEE 31st International
Conference on Advanced Information Networking and Applications (AINA), Taipei, 2017,
pp. 1055-1062.