2018 IEEE Latin American Conference on Computational Intelligence (LA-CCI) Guadalajara, Jalisco, Mexico, November 7-9, 2018
A comparative between Mel Frequency Cepstral
Coefficients (MFCC) and Inverse Mel Frequency
Cepstral Coefficients (IMFCC) features for an
Automatic Bird Species Recognition System
Angel David Pedroza Ramirez
Unidad Academica de Ingenieria Electrica
Universidad Autonoma de Zacatecas
Zacatecas, Mexico
P.A.D 16@hotmail.com
Rogelio Rosas Valdez
Unidad Academica de Ciencias Biologicas
Universidad Autonoma de Zacatecas
Zacatecas, Mexico
rogrosas@gmail.com
Abstract—In this paper a comparative between Mel Fre-
quency Cepstral Coefficients (MFCC) and Inverse Mel Frequency
Cepstral Coefficients (IMFCC) features for an automatic bird
species recognition system is proposed with the aim to validate
IMFCC as a feature that can also be extracted for bird species
recognition. In biodiversity monitoring task there are some
traditional techniques and, bioacoustics studies biodiversity by a
noninvasive way based on the relationship between animal species
and its sounds. Bioacoustics methodology for avian conservation
are based on automatic speech recognition techniques and one
of the traditional extracted features in this area are MFCC.
Nevertheless some new studies uses IMFCC as a complementary
frequency information. From results, it is concluded that IMFCC
features have better performance than traditional MFCC features
but, performance still depends on the recognized bird sound.
Index Terms—bioacoustics, bird classification, HMM, IMFCC,
MFCC.
I. I NTRODUCTION
Biodiversity monitoring represent an invaluable tool in
conservation and climate change reduction [1], [2]. Although
there are some traditional invasive monitoring techniques, a
noninvasive method is preferred. Bioacoustics studies biodi-
versity based on the relationship between animal species and
its sounds.
Bird species are very sensitive to climate change and
its conservation and understanding represents and important
challenge [3]. In this sense, bioacoustics traditional method-
ology for avian conservation are based on automatic speech
The authors would like to thank CONACYT and to the Programa de
Doctorado en Ciencias de la Ingenieria from the Universidad Autonoma de
Zacatecas.
978-1-5386-4626-7/18/$31.00 © 2018 IEEE
Jose Ismael de la Rosa Vargas
Unidad Academica de Ingenieria Electrica
Universidad Autonoma de Zacatecas
Zacatecas, Mexico
ismaelrv@yahoo.com
Aldonso Becerra
Unidad Academica de Ingenieria Electrica
Universidad Autonoma de Zacatecas
Zacatecas, Mexico
a7donso@hotmail.com
recognition techniques [4], [5], [6], [7]. Although there are a
huge variability of techniques and features to be extracted from
sound, one of the traditional features in speech recognition,
are Mel frequency cepstral coefficients (MFCC). Despite the
high efficiency by using MFCC features, some new studies
proposes the use of inverse mel frequency cepstral coefficients
(IMFCC) as a new perspective that take into account frequency
information missed by MFCC features improving the spectral
modeling [8], [9], [10], [11].
Even though MFCC features have been applied in bioacous-
tics, the application of IMFCC features in avian conservation
systems have not been applied yet. In this paper a comparative
between MFCC and IMFCC features for an automatic bird
species recognition system is conducted with the aim to
validate the use of IMFCC features as a feature that can also be
extracted for bird species recognition. From results, it can be
observed that IMFCC features give better performance than
traditional MFCC features. However, the performance still
depends on the bird species to be recognized.
In the following: Section II explains basic mathematical
background of MFCC and IMFCC features. Section III present
the used automatic bird species recognition system. Section
IV summarized the comparative results between MFCC and
IMFCC features, and some discussion. Finally, Section V
concludes this paper.
II. F EATURE E XTRACTION : MFCC AND IMFCC
A. MFCC
Human perception of sound can be described by the Mel
scale and MFCCs are based on this configuration. The basic
idea is to capture sound frequencies by a filter bank and
determine its response according to a certain filter gain (which
resolution is high in low frequencies and low for high frequen-
cies) [5], [12].
To compute MFCC features, once bird sound signal (y(n))
is framed and windowed (using a hamming window), for each
frame, power spectrum (y(k))is calculated according to:
Ns
1 X
y(k) = | y(n)WNkn |2 , (1)
N s n=1 Fig. 1. MFCC filter bank.
where WN = exp(−j2π/N s), N s is the number of points
of the 1D Discrete Fourier Transform (1D DFT) and, 1 ≤ k ≤
N s. Next, traditionally set of p Mel spaced triangular filters are
constructed (see Fig. 1). Then, output of filterbank (E(j)) is
calculated by multiplying each frame energy spectrum (y(k))
and filters gain (Mj (k)) by:
N s/2−1
X
E(j) = y(k)Mj (k), (2)
k=1 Fig. 2. IMFCC filter bank.
where j is the index of each filter in the filter bank and,0 ≤
j < p.
After that, Discrete Cosine Transform (DCT-II) is calculated III. B IRD SPECIES RECOGNITION ALGORITHM
according to:
The basic idea behind the proposed algorithm is that, by an-
p−1
X 
  1
π  alyzing bird sounds (songs and calls), identify an specific bird
Cl (t) = log E j cos t j − , (3) species among the huge variability of bird species. This task
j=0
2 p
represents a big challenge especially under hard environment
where Cl (t) is the tth order MFCC of the lth frame and, record conditions. From this perspective, a lot of techniques
t=1,...,13. Finally, mean value of each frame coeficientes is and automatic bird species recognition configuration system
then calculated (this process is repeated per bird sound audio where develop to complete this task.
file). In this section an automatic bird species recognition system
B. IMFCC used to perform the comparatives is described (see Fig. 3).
On the other hand, to capture the acoustic information A. Pre-processing
missed by the MFCC features, Inverted Mel Frequency Cep-
stral Coefficients (IMFCC) features are traditionally calcu- 1) “Most important” region extraction and noise elimi-
lated. In other words, IMFCC filter bank is configured to nation.: The proposed step is presented as a new “visual
have more frequency resolution in high frequencies and a low filter” technique that, by the user observation-decision of two
resolution for low frequencies [8], [9]. points from the spectrogram, extract the “most important”
To compute IMFCC features, once power spectrum (y(k)) information and/or eliminate a “noise region”. Let f (n) be
is calculated, a set of p Imel spaced triangular filters are a N-point 1D sequences (with index range n = 0, . . . , N − 1).
constructed (see Fig. 2). Next, the output of the IMFCC filter Then, from a user spectrogram selected “most important”
bank (E 0 (j)) is calculated according to: region, 1D DFT of the sequence (F (k)) is calculated and
frequency selected index from region are mapped into 1D DFT
N s/2−1
X coordinates. Then, the minimum of 1D DFT is replaced in the
E 0 (j) = y(k)Mj0 (k), (4) magnitude value of the frequencies that did not correspond
k=1
to the selected region. After that, calculate the 1D Inverse
where Mj0 (k) is the Imel filters gain. Discrete Fourier Transform (1D IDFT) from signal according
Then, Discrete Cosine Transform (DCT-II) is calculated to:
according to:
p−1 N −1
1 π 1 X
f 0 (n) = F (k)WN−kn ,
X
Cl0 (t) = 0
log[E (j)] cos[t(j − ) ], (5) (6)
2 p N
j=0 k=0

where Cl0 (t) is the tth order IMFCC of the lth frame and, where f 0 (n) is the frequency region of interest of the bird
t=1,...,13. Finally, as in MFCC features, mean value of each sound signal. After that, time domain region of interest are
frame coeficientes is then calculated (process is repeated per mapped and then extracted into time domain coordinates.
bird sound audio file). Finally, after normalization, based on an audio playback, the
process is repeated until the user decides that the resulting
audio record has the desired “most important” information.
Fig. 3. Bird sound recognition algorithm.
On the other hand, a “noise region” is eliminated by the
user selection of two points in the spectrogram. In this sense,
as in previous step, 1D DFT from region (G(k)) is calculated,
spectrogram selected frequency information (from the “noise
region”) is replaced with the minimum of 1D DFT magnitude
value and, IDFT (g 0 (n)) is calculated. Finally, resulting region
(g 0 (n)) is replaced in f 0 (n). Also, the process is repeated until
the user decides that the resulting audio record has an adequate
noise elimination.
2) Spectrum threshold: A spectrum threshold is proposed
since not all the information in the spectrogram is important to
0
PN s After f (n) is framed and windowed, for each
be processed.
frame ( n=1 y(n)), energy spectrum (y(k)) is calculated by:
Ns
1 X
y(k) = | y(n)WNkn |2 , (7)
N s n=1
where WN = exp(−j2π/N s) and, 1 ≤ k ≤ N s. After
normalization, an “adequate” magnitud threshold value (thr)
is calculated and for each frame, frequencies magnitud values
in y(k) under thr are set to a minimum value. The optimal
spectrum threshold value in the experiments was a constant
value of thr=0.01.
B. Bird sound classification
Once features are extracted from bird sound signals, tradi-
tional Hidden Markov Models (HMM) are used for training
and testing [13], [?]. Parameters of a model based on the
extracted features (MFCC or IMFCC) are calculated in the
training phase according to a vector of probabilities of each
state in sequence (π), a probability matrix of the transitions
between the states (A) and, a probability matrix for the
emissions given the states (B). This parameters define a bird
sound species model (λ) according to:
λ = (A, B, π), (8)
Baum-Welch algorithm is used for training the previous pa-
rameter set by an iterative process to obtain A,B and π
optimized parameters. In other words, for each bird-sound ex-
tracted features from species it is calculated a model. Finally,
given a bird sound in test step, it is calculated the probability
that the given sound has been produced by a specific bird
species(λ). HMM method was implemented by using a free
software [14].
IV. R ESULTS AND D ISCUSSION
To compare efficiencies of MFCC and IMFCC features by
the proposed algorithm, bird sound files were collected from
the next free online bird sound databases:
• INECOL [15]: This bird sound database is offered as
a toolkit for ornithology research and was constructed
with the aim to study and disseminate the bird sound
recordings in Mexico.
• xeno-canto [16]: Bird sound recording database from a
huge variability of bird species from around the world.
Since xeno-canto recordings are volunteer submitted there
is no uniform recording conditions and, since there are many
types of sounds that can be produced by bird species, same
type recordings was collected from database to create models
from specific bird sounds. Also, INECOL bird sound database
is nowadays limited and therefore only a partial selection of
bird species recordings was made by selecting the bird species
whose type of bird sound has more records in xeno-canto
database (see table I).
Evaluation of the comparatives was made by the following
parameters:
• True Acceptance Rate (TAR): Probability that a bird
sound is correctly identified,
• True Rejection Rate (TRR): Probability that a bird sound
is correctly rejected.
The training set for TAR test was the 70% of bird sound
records per bird species and, for TRR test, it was used the
total of records from others (different to the in acceptance
bird sound) bird species. Results of comparatives are shown
in table II.
As expected, table II shows that the efficiency of the
algorithms change depending on the extracted features. First,
for Automolus rubiginosus it is shown an improvement in TAR
performace and TRR performance by using IMFCC features
(36.37% and 8.69% respectively). Then, for Synallaxis ery-
throthorax it is shown an equal TAR performance for both
techniques but an improvement in TRR performance by using
IMFCC features. Also, for Cardinalis cardinalis it is shown an
improvement in TAR performance (7.69%) by using IMFCC
features but an improvement in TRR performance (23.80%) by
using MFCC features. On the other hand, a random recognition
efficiency is shown in Cercomacra tyrannina.
From general efficiencies obtained in Table II, in contrast
with traditional efficiencies obtained in automatic speech
recognition by using MFCC and IMFCC features (that usually
assumes the use of a large-scale training datasets), it is shown
that the amount of data at the training step plays a significant
role. However, due to the high variability of bird sounds
among all the bird species makes difficult the use of large-
scale training datasets (and therefore limited data is an usual
condition). In addition, the complexity of the bird sounds to
be recognized have an effect in the efficiency of the algorithm.
Therefore depending on the bird sound to be recognized is the
efficiency of the feature extracted (see Fig. 4).
 TABLE I
B IRD SOUND FILES IN COLLECTED DATABASE

Order Family Species Bird sound

files
Passeriformes − Furnariidae Automolus 35
Perching Birds rubiginosus
Synallaxis 54
erythrotho-
rax
Cardinalidae Cardinalis 41
cardinalis (a)
Thamnophilidae Cercomacra 15
tyrannina
Tyrannidae Myiozetetes 33
similis

TABLE II
R ESULTS OF BIRD IDENTIFICATION TEST (%).

MFCC IMFCC
Bird species TAR TRR TAR TRR
Automolus rubiginosus 63.63 34.78 100 43.47 (b)
Synallaxis erythrothorax 100 28.57 100 61.90
Fig. 4. Bird sound spectrogram examples: a) Cercomacra tyrannina song and
Cardinalis cardinalis 76.92 23.80 84.61 0 b) Myiozetetes similis call.
Cercomacra tyrannina 0 0 0 0
Myiozetetes similis 30 4.1 10 61.90

V. C ONCLUSION
In this paper a comparative between MFCC and IMFCC
features for automatic bird species recognition was proposed.
Although both features are traditionally extracted in speech
recognition systems, it is shown that IMFCC features can
also be applied in a bird species recognition system. From
results, IMFCC features have a better general efficiency (in
contrast with traditional MFCC features) but it still depends
on the bird species to be recognized. In the same manner from
results, the “most important” by-observation region extraction
play an important role in the sense that it is expected a better
performance when bird sounds are easy to extract from the
spectrogram.
In future work the use of a traditional fusion feature model
(MFCC-IMFCC) is going to be applied for an automatic bird
species recognition system.
R EFERENCES
[1] C. Bellard, C. Bertelsmeier, P. Leadley, W. Thuiller, and F. Courchamp,
“Impacts of climate change on the future of biodiversity,” Ecology
letters, vol. 15, no. 4, pp. 365–377, 2012.
[2] C. Duncan, J. R. Thompson, and N. Pettorelli, “The quest for a mech-
anistic understanding of biodiversity–ecosystem services relationships,”
Proc. R. Soc. B, vol. 282, no. 1817, p. 20151348, 2015.
[3] J. Wormworth and K. Mallon, “Bird species and climate change: the
global status report-a synthesis of current scientific understanding of
anthropogenic climate change impacts on global bird species now, and
projected future effects,” 2006.
[4] P. C. Caycedo-Rosales, J. F. Ruiz-Muñoz, and M. Orozco-Alzate,
“Reconocimiento automatizado de señales bioacústicas: una revisión de
métodos y aplicaciones,” Ingenierı́a y Ciencia, vol. 9, no. 18, 2013.
[5] C. Chou, P. Liu, and B. Cai, “On the studies of syllable segmentation
and improving mfccs for automatic birdsong recognition,” in 2008 IEEE
Asia-Pacific Services Computing Conference, 2008, pp. 745–750.
[6] T. S. Brandes, “Feature vector selection and use with hidden markov
models to identify frequency-modulated bioacoustic signals amidst
noise,” IEEE Transactions on Audio, Speech, and Language Processing,
vol. 16, no. 6, pp. 1173–1180, 2008.
[7] T. M. Aide, C. Corrada-Bravo, M. Campos-Cerqueira, C. Milan,
G. Vega, and R. Alvarez, “Real-time bioacoustics monitoring and
automated species identification,” PeerJ, vol. 1, p. e103, 2013.
[8] S. Chakroborty and G. Saha, “Improved text-independent speaker identi-
fication using fused mfcc & imfcc feature sets based on gaussian filter,”
International Journal of Signal Processing, vol. 5, no. 1, pp. 11–19,
2009.
[9] D. Sharma and I. Ali, “A modified mfcc feature extraction technique
for robust speaker recognition,” in 2015 International Conference on
Advances in Computing, Communications and Informatics (ICACCI),
2015, pp. 1052–1057.
[10] N. Sen, T. Basu, and S. Chakroborty, “Comparison of features extracted
using time-frequency and frequency-time analysis approach for text-
independent speaker identification,” in 2011 National Conference on
Communications (NCC), 2011, pp. 1–5.
[11] S. Memon, M. Lech, N. Maddage, and L. He, “Application of the vector
quantization methods and the fused mfcc-imfcc features in the gmm
based speaker recognition,” in Recent Advances in Signal Processing.
IntechOpen, 2009.
[12] C. Kwan, G. Mei, X. Zhao, Z. Ren, R. Xu, V. Stanford, C. Rochet,
J. Aube, and K. C. Ho, “Bird classification algorithms: theory and exper-
imental results,” in 2004 IEEE International Conference on Acoustics,
Speech, and Signal Processing, vol. 5, 2004, pp. V–289.
[13] L. Rabiner and B.-H. Juang, Fundamentals of Speech Recognition.
Prentice-Hall, Inc., 1993.
[14] K. Murphy, “Hidden markov model (hmm) toolbox for matlab,”
https://www.cs.ubc.ca/ murphyk/Software/HMM/hmm.html, 1998.
[15] “Inecol,” http://www1.inecol.edu.mx/sonidos/menu.htm, Accessed 16
Marzo 2017.
[16] “Xeno-canto,” https://www.xeno-canto.org, Accessed 07 Febrero 2017.