See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322336038

Automatic Speaker, Age-group and Gender Identification from Children’s

Speech

Article  in  Computer Speech & Language · January 2018

DOI: 10.1016/j.csl.2018.01.001

CITATIONS READS

2 177

3 authors:

Saeid Safavi Martin J Russell

University of Surrey University of Birmingham
27 PUBLICATIONS   131 CITATIONS    134 PUBLICATIONS   1,735 CITATIONS   

SEE PROFILE SEE PROFILE

Peter Jancovic
University of Birmingham
90 PUBLICATIONS   648 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Speech Recognition by Synthesis (SRbS) View project

FSR ground clutter View project

All content following this page was uploaded by Saeid Safavi on 13 February 2018.

The user has requested enhancement of the downloaded file.

 Available online at www.sciencedirect.com

Computer Speech & Language 50 (2018) 141
156

www.elsevier.com/locate/csl

Automatic speaker, age-group and gender identification from

children’s speechI
TagedPSaeid Safavi, Martin Russell*, Peter Jan
covi
c
TagedPDepartment of Electronic, Electrical and Systems Engineering, School of Engineering, University of Birmingham, Birmingham B15 2TT, UK
Received 1 May 2016; received in revised form 22 November 2017; accepted 3 January 2018
Available online 9 January 2018

TagedPAbstract
A speech signal contains important paralinguistic information, such as the identity, age, gender, language, accent, and the emo-
tional state of the speaker. Automatic recognition of these types of information in adults’ speech has received considerable atten-
tion, however there has been little work on children’s speech. This paper focuses on speaker, gender, and age-group recognition
from children’s speech. The performances of several classification methods are compared, including Gaussian Mixture Mod-
el
Universal Background Model (GMM
UBM), GMM
Support Vector Machine (GMM
SVM) and i-vector based
approaches. For speaker recognition, error rate decreases as age increases, as one might expect. However for gender and age-
group recognition the effect of age is more complex due mainly to consequences of the onset of puberty. Finally, the utility of dif-
ferent frequency bands for speaker, age-group and gender recognition from children’s speech is assessed.
Ó 2018 Elsevier Ltd. All rights reserved.

TagedPKeywords: Speaker recognition; Gender identification; Age-group identification; Children’s speech; Gaussian Mixture Model (GMM); Support
Vector Machine (SVM); i-vector

1. Introduction

TagedPSpeech signals contain significant information in addition to their linguistic content, for example information
about the speaker’s identity, gender, social group, geographical origin, health, smoking habit, height and emotional
state. This type of information is referred to as paralinguistic. While speech recognition is concerned with extracting
the underlying linguistic message in a speech signal, paralinguistic analysis is concerned with extracting its paralin-
guistic information.
TagedPFor a subset of paralinguistic technologies, including speaker, language, accent and dialect recognition, a set of
common tools has been developed. Analysis typically begins by transforming a speech signal into a sequence of
spectrum-based feature vectors, for example mel frequency cepstral coefficients (MFCCs) (Davis and Mermelstein,
1980). In the classification stage it is useful to distinguish between acoustic approaches, which assume that a particu-
lar class is distinguished by its distribution of acoustic feature vectors, and phonotactic approaches, which exploit

I
This paper has been recommended for acceptance by Roger Moore
* Corresponding author.
E-mail address: m.j.russell@bham.ac.uk (M. Russell).

http://dx.doi.org/10.1016/j.csl.2018.01.001
0885-2308/ 2018 Elsevier Ltd. All rights reserved.
142 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

TagedPhigher-level sequential structure (Schuller et al., 2013). Phonotactic and acoustic methods provide complementary
information, which state-of-the-art systems exploit by fusing their output scores (Wong and Sridharan, 2001) A
more detailed review of these methods is given in Section 2.
TagedPAlthough there has been considerable work on paralinguistic analysis of adults’ speech (Schuller and Batliner,
2013; Schuller et al., 2013), there is relatively little work on this topic for children’s speech. The most notable excep-
tion is the 2009 paralinguistic challenge (Schuller et al., 2011) held at Interspeech 2009, which focused on emotion
recognition for the German FAU Aibo Emotion Corpus (Batliner et al., 2004) of spontaneous speech from 10 to 13
year old German children interacting with an Aibo robotic dog. Another example is early work on automatic assess-
ment of goodness of pronunciation in children’s speech (Russell et al., 2000).
TagedPThe paucity of work in this area on children’s speech is surprising because paralinguistic analysis has a number of
significant applications that are especially relevant for children. For example, voice-based speaker, gender and age
recognition offer novel approaches to safeguarding children engaged in social networking, gaming or general inter-
net browsing. Also, speaker recognition provides an alternative means to identify which child is using a particular
piece of educational software or whether a child is present in a school or class.
TagedPChildren’s speech poses particular problems. It is well documented (for example Lee et al., 1997; Lee et al., 1999;
Gerosa et al., 2006; Gerosa et al., 2007) that spectral and durational variability are greater for children’s speech than
for adults’ speech, converging to adult levels at around the age of 13 years. However, it is not known whether this is
mainly due to developing motor control skills or cognitive issues relating to language acquisition. Because of their
shorter vocal tracts, important structure in speech occurs at higher frequencies for children than for adults. Conse-
quently the effects of bandwidth reduction on computer and human speech recognition performance are much more
pronounced for children’s speech (Russell et al., 2007). In addition, children’s higher fundamental frequencies can
result in greater variability in spectrum-based feature vectors, such as MFCCs, used in automatic speech recognition
(ASR) (Ghai, 2011).
TagedPUnlike paralinguistic analysis, ASR for children’s speech has been an active research topic for the past 20 years.
Due to the factors cited above, recognition error rates are typically much greater for children’s speech, especially for
young children, even if the recogniser is trained on age-matched data (Wilpon and Jacobsen, 1996). The availability
of ASR systems trained on adult speech and the absence of large corpora of children’s speech, has motivated
research into the use of adaptation techniques such as vocal tract length normalisation (VTLN) (Lee and Rose,
1998) and pitch normalisation (Ghai, 2011), as well as generic adaptation techniques such as MAP and MLLR adap-
tation (Elenius and Blomberg, 2005; Gerosa et al., 2005), to enable adult ASR systems to recognise children’s
speech. In the most recent research, deep neural network (DNN) based ASR has been applied to children’s speech
(Liao et al., 2015; Serizel and Giuliani, 2017).
TagedPIn addition to the above, another significant factor that complicates the identification of age-group and gender for
older children is the onset of puberty. Rogol et al. (2000) observed wide variations in the timing of the onset of
puberty and in the consequent physiologic variations in growth. They estimate the age of the onset of puberty for
girls and boys as 11 and 13 years, respectively.
TagedPThe objectives of the work described in this paper are to establish baselines for paralinguistic analysis of child-
ren’s speech, focussing on the application of state-of-the-art methods to the identification of a child’s identity, gen-
der, and age-group, and to investigate the importance of information in different frequency sub-bands for these
tasks.
TagedPThe paper is organised as follows. Section 2 presents a review of approaches to paralinguistic analysis of adult
speech, including speaker, language, regional accent and dialect recognition. Section 3 describes the data that was
used and how it was partitioned for speaker, gender and age-group recognition. This section also describes feature
extraction and the classification methods that were used. Section 4 investigates the utility of different frequency sub-
bands for these technologies. Section 5 presents the results of speaker, gender and age-group recognition experi-
ments with ‘full bandwidth’ children’s speech, and Section 6 presents our conclusions.

2. Paralinguistic analysis of adult speech

TagedPThe simplest acoustic approach to speaker and language recognition for adults’ speech is to model the distribution
of feature vectors for a particular class as a Gaussian Mixture Model (GMM) (Reynolds and Rose, 1995; Naik et al.,
1989). First, a class-independent model, λUBM called the Universal Background Model (UBM) is estimated using all
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 143

TagedPof the available training data. A class-dependent GMM λc is then created for each class c from the UBM using
training data for c and MAP (Gauvain and Lee, 1994) adaptation. In the GMM
UBM approach, classification
of an utterance u is based on the probabilities p(ujλc). In speaker and language identification this probability
is used directly, whereas for speaker verification p(ujλs) is divided by p(ujλUBM), and the result is compared
with a threshold. Score normalisation techniques, for example T-norm (Auckenthaler et al., 2000), Z-norm
(Wan and Campbell, 2000), and max-norm (Campbell and Karam, 2009) enable the same threshold to be used
for all speakers.
TagedPThe GMM
UBM approach has been superseded by methods that map an utterance u into a vector space and then
apply a static classifier such as a Support Vector Machine (SVM) (Campbell et al., 2006). For example, MAP adapta-
tion of the UBM is performed using u, and the mean vectors of the components of the resulting GMM are
concatenated to form a GMM ‘supervector’ (Campbell et al., 2006). Inter-session variability, including inter-channel
variability or channel factors, is a significant source of error for these techniques. Intersession variability compensa-
tion (ISVC) (Vair et al., 2006) and nuisance attribute projection (NAP) (Solomonoff et al., 2004) address this issue
by isolating this variability in a linear subspace of the supervector space. In more recent approaches based on Joint
Factor Analysis (JFA) (Kenny et al., 2007; Dehak et al., 2011a), and i-vectors (Dehak et al., 2011b), the GMM super-
vector is mapped into a lower-dimensional discriminative total variability space. These methods have been applied
to a number of problems, including identification of speaker (Campbell et al., 2009; Kinnunen and Li, 2010), lan-
guage, regional accent and dialect (Biadsy et al., 2011; DeMarco and Cox, 2012; Hanani et al., 2013), and age and
gender (Metze et al., 2007; Bocklet et al., 2008; Dobry et al., 2009; Mahmoodi et al., 2011; Chen et al., 2011;
Van Heerden et al., 2010; Bahari and Van Hamme, 2011; Bahari et al., 2012).
TagedPDifferent frequency regions contain different types of linguistic and paralinguistic information. For instance, fre-
quency regions below 600 Hz and above 3000 Hz are the most informative for speaker identification from adults’
speech (Besacier et al., 2000; Safavi et al., 2012; Orman and Arslan, 2001).
TagedPIn phonotactic approaches, a spoken utterance is transformed into a sequence of tokens from a finite set,
and it is assumed that different classes are distinguishable from the statistics of these sequences. In the Phone
Recognition followed by Language Modelling (PRLM) and Parallel PRLM (PPRLM) approaches to language
identification of Zissman (1996), an automatic phone recognition system transforms the utterance into a
sequence of phones, which is analysed using language-dependent statistical language models. In Hanani et al.
(2013) the tokens are GMM component indices and the component n-gram statistics are collated in a vector
which is classified using language-dependent SVMs. Similar methods are employed in the ‘ideolect’ approach
to speaker verification (Doddington, 2001).
TagedPRecently, deep neural networks (DNNs) have been employed for speaker (Yamada et al., 2013; Lei et al., 2014;
Variani et al., 2014; Sarkar et al., 2014) and language recognition (Song et al., 2013; Lopez-Moreno et al., 2014).
One approach is to use a DNN directly as a classifier to discriminate between classes, with the frame-level DNN pos-
teriors being averaged over the duration of the utterance (Lei et al., 2014). An alternative is to use a DNN to extract
bottleneck features or posteriors which are then used to train an i-vector classifier (Song et al., 2013; Sarkar et al.,
2014).

3. Method

3.1. Speech data

TagedPThe OGI Kids Speech corpus (Shobaki et al., 2000) is used in all experiments. It contains examples of spontane-
ous and read speech from approximately 1100 children (roughly 100 children per grade from kindergarten
(5
6 years old) to grade 10 (15
16 years old)), recorded in Portland, Oregon. Prompts were displayed as text on a
screen, and a human recording of the prompt was played in synchrony with facial animation using the animated 3D
character Baldi (Massaro, 1998). The subject repeated the prompt (read speech) or talked on a chosen topic (sponta-
neous speech). Recordings were made at 16 kHz sampling rate, 16 bits precision, using a head-mounted microphone.
An average session lasted 20 min and yielded approximately 63 utterances (8
10 min of speech). However, for some
subjects much less data is available. In speaker recognition experiments just 144 and 48 s of speech were used per
subject for training and testing, respectively, for consistency between subjects. The numbers, ages and genders of
the children recorded in the corpus are shown in Table 1.
144 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

Table 1
Age-range, gender split and number of children recorded per grade in the OGI Kids Speech corpus.
Grade K 1 2 3 4 5 6 7 8 9 10
Age 5
6 6
7 7
8 8
9 9
10 10
11 11
12 12
13 13
14 14
15 15
16

Male 39 58 53 63 47 49 57 46 49 70 76
Fem. 50 31 61 54 45 49 55 51 50 40 30
Total 89 89 114 117 92 98 112 97 99 110 106

3.2. Partitioning of the data

TagedPUnderstanding how the performance of paralinguistic technologies varies with age is another objective of this
work. Because only a relatively small amount of data is available for each grade, grades are combined into three
age-groups, referred to as AG1, AG2 and AG3. The grade that is likely to differ most from the others is K, because
the speech of 5
6 year-olds may exhibit additional variability due to phonological factors associated with language
acquisition (Fringi et al., 2015). Hence the first age-group, AG1, should span as few grades as possible. For speaker
recognition, AG1 comprises grades K, 1 and 2. However, because there is relatively little speech from some subjects
in grades K, 1 and 2, for gender and age-group identification grade 3 was included in AG1 and the boundaries for
AG2 and AG3 were shifted by one grade. Table 2 summarises the allocation of grades to age-group for the different
experiments.
TagedPDifferent experiments require different partitions of the corpus into training and test sets, to achieve a suitable bal-
ance between classes or to study a particular effect. For speaker recognition, different configurations are investi-
gated, but in general 144 s of speech per speaker are used for training, and segments of length 10 s are used for
testing. For gender identification the test set comprises utterances from 687 subjects, each of length 30 s, balanced
by age and gender, and all recordings from the remaining 413 speakers are used for training. For the age-group iden-
tification task a gender balanced test set of recordings from 766 speakers (290, 285 and 191 from AG1, AG2 and
AG3, respectively), each of length 30 s, is used. The recordings from the remaining 344 speakers were used for train-
ing. Details are included in the descriptions of the individual experiments. For building the UBM all the training data
available in the corpus is used and we have not used any development set. This is summarised in Table 3.

Table 2
Definitions of the age-groups used in the
experiments.
Experiment AG1 AG2 AG3

Speaker recognition K
2 3
6 7
10

Gender identification K
3 4
7 8
10
Age-group identification K
3 4
7 8
10

Table 3
Details of training and test data used in the experiments. s/speaker indicates the approximate duration of
speech (in seconds) per speaker. FB, SB(a) and SB(b) stand for full-band, gender-independent sub-band and
gender-dependent sub-band experiments, respectively.
Training set Test set

Num. of Duration Num. of Duration

speakers (s/speaker) speakers (s/speaker)

Speaker rec. (FB) 864 144 864 10

Speaker rec. (SB(a)) 359 144 359 10
Speaker rec. (SB(b)) 864 144 864 10
Speaker rec. (Class) 30 144 30 10
Gender ID (FB) 413 All 687 30
Gender ID (SB) 413 All 687 30
Age-group ID (FB) 344 All 766 30
Age-group ID (SB) 344 All 766 30
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 145

Table 4
The centre frequencies for 24 Mel-spaced critical-band filters.
Filter number 1 2 3 4 5 6 7 8
Centre frequency 156 281 406 500 625 750 875 1000
Filter number 9 10 11 12 13 14 15 16
Centre frequency 1125 1281 1437 1625 1843 2062 2343 2656
Filter number 17 18 19 20 21 22 23 24
Centre frequency 3000 3375 3812 4312 4906 5531 6281 7093

3.3. Signal analysis

TagedPSilence was discarded using energy-based speech activity detection. Frames of length 20 ms (10 ms overlap,
Hamming window) were extended to 512 samples and a DFT was applied, giving a frequency resolution of
31.25 Hz. The resulting magnitude spectrum was passed to a bank of 24 Mel-spaced triangular filters, spanning fre-
quencies from 0 Hz to 8000 Hz. Table 4 shows the centre frequency of each filter, quantised to the nearest 31.25 Hz
(the cut-off frequencies of a filter are the centre frequencies of the adjacent filters). For the full bandwidth experi-
ments the outputs of all 24 filters were transformed into 19 static plus 19 D and 19 D2 MFCCs.

3.4. Classification

TagedP3.4.1. Automatic classification

TagedPThree types of classifier are considered, GMM
UBM (Reynolds et al., 2000; Campbell et al., 2006),
GMM
SVM (Campbell et al., 2006) and i-vector (Singer et al., 2012; Dehak et al., 2011b) based methods. All iden-
tification and verification experiments were text-independent.
TagedPIn the GMM
UBM approach (Reynolds et al., 2000), class-dependent GMMs are obtained by MAP adaptation
(Gauvain and Lee, 1994) of the component mean vectors of the UBM (“class” is speaker, gender or age-group),
using 144 s of data per speaker for speaker recognition, and all of the training data for gender and age-group identifi-
cation. Speaker verification followed the NIST methodology (NIST, 2003), whereby a test utterance is compared
with eleven models, the “true” model plus ten randomly selected impostor models. For age-group identification, gen-
der-dependent and gender-independent models are created. For speaker verification the scores were normalised using
“max-norm” (Campbell and Karam, 2009).
TagedPIn the GMM
SVM system, the mean vectors of class-specific GMMs were concatenated into supervectors
(Campbell et al., 2006). These supervectors were used to build one SVM per class, treating that class as the ‘target’
and the others as ‘background’. For age and gender recognition, one supervector was created per speaker. In speaker
recognition, the training data was split into nine 16 s segments to create 9 supervectors per speaker.
TagedPThe third approach uses i-vectors. This assumes that a GMM supervector m can be decomposed as m ¼ m þ Tw;
where m is the UBM mean supervector, T is a linear mapping from a low-dimensional ‘total variability subspace’ W
into the supervector space, and w 2 W is an ‘i-vector’, sampled from a standard normal distribution. Two approaches
to i-vector scoring were employed. In the first (Singer et al., 2012), linear discriminant analysis (LDA) is applied to
W, and each class c is represented by the mean mc of length normalised i-vectors for that class in the LDA sub-space.
At the recognition stage, the score for each class c is the dot product of the normalised LDA test i-vector with mc (the
cosine of angle between the test i-vector and mc). This approach is referred to as ‘i-vector’. The second i-vector
method, applied to speaker recognition, uses probabilistic LDA (PLDA) (Prince and Elder, 2007). Before applying
PLDA, the i-vectors are length-normalised and whitened (Garcia-Romero and Espy-Wilson, 2011). Recognition is
based on the log-likelihood ratio between same versus different speaker hypotheses (Garcia-Romero and Espy-Wil-
son, 2011). This is referred to as ‘i-vector-PLDA’. In all experiments, the size of the T matrix and the i-vector dimen-
sion after LDA were set, empirically, to 400 and 300, respectively. The T matrix was learned using all training data.
TagedPIn NIST evaluations (NIST, 2003) test utterances are 10, 20 or 30 s long. We used 10 s test utterances for speaker
identification and verification and 30 s utterances for age and gender identification.
146 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

TagedP3.4.2. Experiments with human listeners

TagedPHuman gender and age-group identification performance for children’s speech was also measured, using test
utterances from the automatic classification experiments. Twenty listeners, mainly research students, participated in
the evaluations. Since few of these had children they cannot be considered to be “expert” listeners. Each participant
listened to an average of 34 and 38 utterances, each of duration 10 s, for gender-ID and age-group-ID, respectively,
in a quiet room using the same computer and headphones.

4. Experimental results I: sub-bands

TagedPThe purpose of this section is to identify the most important frequency bands for speaker, gender and age-group
identification from children’s speech. During feature extraction (Section 3.3) the short-term magnitude spectrum is
passed through a bank of 24 Mel-spaced triangular filters, spanning frequencies from 0 Hz to 8000 Hz. Sub-band
classification experiments were conducted using speech data comprising the outputs of groups of 4 adjacent filters.
We considered 21 overlapping sub-bands, where the Nth sub-band comprises the outputs of filters N to N þ 3 (N=1
to 21). Each set of 4 filter outputs was transformed to 4 MFCCs plus 4 D and 4 D2 parameters.
TagedPFeature warping, which improves performance in adult speaker recognition (Pelecanos and Sridharan, 2001), was
applied to transform the short-term distribution of features to be normally distributed, using a 3 s sliding window.
Systems were built for each of the 21 sub-bands using 64-component GMM
UBM systems (the choice of 64 com-
ponents was determined empirically on an evaluation set disjoint from the test set). It will be seen (Section 5) that
GMM
UBM systems do not give best performance in full-band experiments. However, the purpose of the present
study is to measure the relative contribution, rather than optimise performance, of sub-bands for speaker-, age-group-
and gender-identification.

4.1. Speaker recognition

TagedPResults of sub-band speaker identification experiments, using a test set of 10 s recordings from 359 speakers bal-
anced for age and gender, are presented in Fig. 1(a). The sub-band speaker identification accuracies vary between
5% and 34%. A second test set, comprising 288 recordings of duration 10 s each, from each of AG1, AG2 and AG3
(Section 3.1), was used to investigate the effect of age on sub-band recognition accuracy. Fig. 1(b) shows sub-band
speaker identification rates for children in AG1, AG2, and AG3. In most cases, the identification rate is highest for
the older children, and decreases for younger children.

30 30

AG1 (K-2nd grade speakers)

rd th
AG2 (3 -6 grade speakers)
25 25 th th
AG3 (7 -10 grade speakers)

20 20
Identification Rate (%)
Identification Rate(%)

15 15

10 10

5 5

0 0
S1 S5 S9 S13 S17 S21 S1 S5 S9 S13 S17 S21
Sub-Band Sub-Band

(a) (b)
Fig. 1. Speaker identification accuracy for each frequency sub-band for the GMM
UBM system: independent of age (a) and for age-groups AG1,
AG2 and AG3 (b).
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 147

TagedPBy analogy with adult speech (Besacier et al., 2000; Orman and Arslan, 2001; Safavi et al., 2012), it is useful to
partition the spectrum into frequency regions. We consider four regions, B1
B4, comprising sub-bands 1
5
(0
1.13 kHz), 6
14 (0.63
3.8 kHz), 15
18 (2.1
5.53 kHz), and 19
21 (3.4
8 kHz), respectively. The most use-
ful bands for speaker recognition are B1, corresponding to individual differences in primary vocal tract reso-
nances and nasal speech sounds, and B3, corresponding to high-frequency speech sounds. The importance of
fricatives (hence region B3) for speaker recognition was noted in Parris and Carey (1994), where it is hypoth-
esised that this may be due to individual differences in the anatomy of the dental arch. Frequency regions
similar to B1
B4 were identified in Besacier et al. (2000) and Orman and Arslan (2001) for adult speaker
recognition. However, the frequencies at which these bands occur in children’s speech are increased by
approximately 38% (B1), 21% (B2) and 11% (B3) relative to adults. One would expect the bands to occur at
higher frequencies for young children, where formants and other structures will occur at higher frequencies
due to their shorter vocal tracts.

4.2. Gender identification

TagedPFig. 2 shows results, averaged across age-groups, of sub-band gender-ID experiments using age-dependent sub-
band GMM
UBM systems. The test set comprises 687 recordings of child speech, each of duration 30 s, balanced
by age and gender, and all recordings from the remaining speakers were used for training. For most sub-bands, accu-
racy is better than chance. The most useful sub-bands are from 8 to 12 (0.9
2.6 kHz), which corresponds, approxi-
mately, to the location of the second formant. This region also provides the best gender-ID performance for adults
(Wu and Childers, 1991).
TagedPFig. 3 shows gender-ID accuracy per sub-band by age-group. For sub-bands 1
10 and 20
21 there is a strong
relationship between accuracy and age, with AG1 achieving the poorest and AG3 the best results. Indeed, for AG1,
the performance is close to chance for sub-bands up to 7 (1.4 kHz). By contrast, the performance for sub-bands 10,
11 and 12 is comparable to the full bandwidth performance for AG3. For sub-bands 12
19 the picture is less clear
and the variations in performance for the different age-groups is less marked.
TagedPWe speculated that poor performance for sub-bands up to 11 for young children (AG1) might be due to the wide
spacing of their pitch harmonics, compared with the narrow low frequency Mel-scale filters (Ghai, 2011). In auto-
matic speech recognition, this is mitigated by broadening these filters so that the equivalent rectangular bandwidth is
300 Hz (Ghai, 2011). The results of applying this to the filters in sub-bands 1
11 are shown as cyan bars in Fig. 3,
which indicates significantly improved performance.

75
Full−bandwidth performance
72
70

65
Identification Rate (%)

60

55

50

45

40
S1 S5 S9 S13 S17 S21
Sub−Band

Fig. 2. Gender-ID accuracy for different frequency sub-bands averaged over all age-groups obtained by the GMM
UBM system.
148 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156
80
78
76

70
Identification Rate (%)

63

60

AG1 (K-2 nd grade speakers)

50 AG1 (K-2 nd grade speakers) - redesigned filters
AG2 (3 rd-6 th grade speakers)
AG3 (7 th-10th grade speakers)
AG1 - FB
AG2 - FB
AG3 - FB
40
S1 S5 S9 S 13 S 17 S 21
Sub-Band

Fig. 3. Gender-ID accuracy for different frequency sub-bands for each of the three children age-groups obtained by the GMM
UBM system. The
horizontal lines indicate the full-bandwidth (FB) gender-ID accuracy for each age-group. (For interpretation of the references to colour in this
figure, the reader is referred to the web version of this article).

4.3. Age-group identification

TagedPAll sub-band age-group-ID experiments use gender-independent GMM
UBM systems and the test set of 30 s
recordings from 766 speakers (290, 285 and 191 from AG1, AG2 and AG3, respectively), balanced for gender. The
recordings from the remaining 344 speakers were used for training. The use of gender-independent models is justi-
fied in Section 5.3.
TagedPFig. 4(a) presents average age-group-ID accuracy for each sub-band. Even for narrow frequency regions, perfor-
mance is above chance in most cases. The best performance is achieved using sub-bands 13
16, while sub-bands
18
21 are the least useful for age-group-ID.
0.015
70

65
0.01

60
Identification Rate (%)

0.005
NGID-NAID

55

50 0

45

-0.005
40

35 -0.01
S1 S5 S9 S13 S17 S21 S1 S5 S9 S13 S17 S21
Sub−Band Sub-Band

(a) (b)
Fig. 4. Age-group-ID accuracy (a) and the differences between normalised gender-ID and age-group-ID (NGID
NAID) accuracy (b) for each fre-
quency sub-band.
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 149

TagedPFig. 4(b) contrasts the usefulness of sub-bands for age-group-ID and gender-ID. The figure was obtained by nor-
malising the data in Figs. 2 and 4(a) so that in both cases the sum of the values over all of the sub-bands is 1. The nor-
malised age-group-ID results were then subtracted from the normalised gender-ID results to obtain Fig. 4(b) (a
similar procedure is described in Safavi et al., 2012). Positive and negative values in Fig. 4(b) indicate sub-bands
which are more useful for gender-ID and age-group-ID, respectively. The most useful sub-bands for age-group-ID,
relative to gender-ID, are 3 and 4 (281
625 Hz), and from 13 to 16 (1.62
3 kHz). Thus, while gender-ID appears to
make use of similar information to speaker recognition, age-group-ID is more similar to speech recognition or accent
ID (Safavi et al., 2012).

5. Experimental results II: full bandwidth

5.1. Speaker recognition

TagedPTable 5 shows full-bandwidth speaker verification and identification results for age-groups AG1
AG3 (288 chil-
dren per group, 10 s test utterances), using 1024 component GMM
UBM, 64-component GMM
SVM systems and
i-vector-PLDA. The number of GMM components was determined empirically using a separate evaluation set which
is disjoint from the test set.
TagedPIn most cases verification EER decreases and identification accuracy increases with age. For example, for the
GMM
UBM system the EER drops by 70% from 2.1% for AG1 to 0.67% for AG3. The corresponding increase in
identification rate is 38%. The i-vector-PLDA system achieves the best overall performance.
TagedPComparison with results for adult speech is complicated by the focus of current research on conversational tele-
phone speech with challenging channel and noise conditions. However, GMM
UBM speaker identification error
rates less than 0.5% and verification equal error rates of 0.24% have been reported for TIMIT (Reynolds, 1995).
Although these are plausible limiting values for the scores in Table 5, they suggest that even for children in AG3
speaker recognition is more challenging than for adults’ speech.
TagedPTable 6 shows the results of simulated identification and verification of a child in a classroom, using i-vector-
PLDA and 10 s test utterances. The experiment uses 12 “classrooms” of 30 children, 4 for each age-group, balanced
across gender. The results follow the expected pattern, with identification accuracy increasing and verification EER
decreasing with age.

5.2. Gender identification

TagedP5.2.1. Age-independent and age-dependent modelling

TagedPAge-independent GMM
UBM, GMM
SVM and i-vector gender-ID systems were evaluated on 30 s test utter-
ances from 687 speakers balanced by age and gender. In the age-dependent systems a separate system was created
for each age-group. Table 7 shows that the age-dependent systems perform best, with accuracy increasing from
67.39%, for the age-independent GMM
UBM system, to 79.18% for the age-dependent GMM
SVM system. In
both cases the GMM
SVM systems achieve the best results. The numbers of GMM components in the age-

Table 5
Speaker verification equal error rate (EER) and identification
accuracy for three different age-groups of children (AG1, AG2
and AG3).
GMM
UBM GMM
SVM i-vector
(1024) (64) (PLDA)

Verification
EER (%)

AG1 (K-2nd) 2.10 1.80 1.59
AG2 (3rd
6th) 1.33 1.21 0.72
AG3 (7th
10th) 0.67 0.64 0.53
Identification
Acc (%)
AG1 (K-2nd) 62.15 75.00 78.16
AG2 (3rd
6th) 80.56 89.24 91.13
AG3 (7th
10th) 85.71 93.26 95.38
150 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

Table 6
Speaker verification EER and identification
accuracy for three different age-groups (AG1,
AG2 and AG3) obtained using the i-vector-
PLDA system for the case of a child in a class-
room scenario.
Classroom scenario Verification Identification
EER (%) Acc (%)

AG1 (K-2nd) 1.89 86.39

AG2 (3rd
6th) 1.39 93.61
AG3 (7th
10th) 0.97 97.34

Table 7
Gender-ID accuracy using age-independent and age-
dependent GMM
UBM, GMM
SVM, and i-vector
systems.
System Gender-ID rate (%) Gender-ID rate (%)
(age-independent) (age-dependent)

GMM
UBM 67.39 71.76

GMM
SVM 77.44 79.18
i-vector 74.26 72.54

TagedPindependent and -dependent GMM
UBM, GMM
SVM and i-vector systems were 1024, 512, and 256, and 512,
256 and 128, respectively.
TagedPThe application of intersession variability compensation (ISVC) (Vair et al., 2006) resulted in only small
improvements. Specifically, the accuracies of the age-independent and -dependent systems improved from 67.39%
to 69.29%, and 71.76% to 72.81%, respectively.

TagedP5.2.2. Effect of age on gender-identification performance

TagedPTable 8 shows the performance of the age-dependent GMM
SVM system according to child age and gender. For
AG1 and AG2, performance for boys is poorer than girls. The accuracy for AG2 is considerably higher than for
AG1 (the performance increase is 7.34%), but for AG3 the accuracy is unexpectedly low. We speculated that this is
because the boys in AG3 fall into two subsets, according to whether or not their voices have broken as a consequence
of reaching puberty, and boys whose voices have not broken may achieve a better match with the “girls” model. It

Table 8
Gender-ID accuracies (in %) for the age-
dependent GMM
SVM system for boys
(“B”), girls (“G”) and overall (“Av.”), and
overall (“Av.”) accuracy for human listen-
ers, for each age group.
Age GMM
SVM Human

group B G Av. Av.

AG1 73.33 80.86 76.80 60.48

AG2 79.50 88.70 84.14 70.49
AG3 77.69 72.13 75.91 70.90
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 151

Table 9
Confusion matrix for gender identification (in %) for age-
group AG3 when using (a) a single model for boys, and (b)
two separate models for boys, corresponding to broken BB
and unbroken BU voices.
(a) (b)

B G BB BU G

B 71.5 28.5 BB 96.2 3.8 0

G 11.5 88.5 BU 5.8 94.2 0
G 6.5 23.0 70.5

TagedPhas been estimated (Rogol et al., 2000) that puberty starts at a skeletal (biological) age of 11 years in girls and
13 years in boys.
TagedPTable 9(a) is the confusion matrix for AG3 for the GMM
UBM system from Table 8, which uses two gender-
dependent models. It indicates that 28.5% of boys are misclassified as girls, compared with only 11.5% of girls mis-
classified as boys. The focus on GMM
UBMs is due to the limited data available in the next experiment.
TagedPA subset of the boys data in AG3 was split into two separate classes: those whose voices have broken (BB), or
remained unbroken (BU), each containing recordings from 18 speakers. A human listener allocated the boys’ data to
these two classes. This data, together with recordings from 18 of the girls in AG3 (G), was used to create 3
GMM
UBM sub-group models, corresponding to (BB), (BU) and (G). The number of GMM components was
reduced to 128 because of the limited data.
TagedPTable 9(b) shows the confusion matrix for AG3 for the three gender sub-group GMM
UBM systems. The overall
gender identification rate for AG3 is 87.43%, compared with 78.53% using two gender models, and the number of
boys misclassified as girls is reduced to zero. Girls are misclassified as boys with unbroken voices 3.5 times more
often than as boys with broken voices. Gender-ID accuracy for girls decreases, because more girls are confused with
boys with unbroken voices, suggesting that an appropriate partition of the girls’ data may also improve performance.

TagedP5.2.3. Human performance

TagedPThe final column of Table 8 shows gender-ID accuracies achieved by human listeners. The average performance
over all age-groups is 66.96%. Evidently humans, as well as computers, have difficulty identifying the gender of a
child from his or her speech.

5.3. Age-group identification

TagedP5.3.1. Gender-independent and -dependent modelling

TagedPTable 10 presents age-group-ID results for GMM
UBM, GMM
SVM, and i-vector systems. The test set is the
same set of 766 recordings, each of length 30 s, that was used in the sub-band age-group-ID experiments. The first
two rows of the table compare the age-group-ID accuracies obtained using gender-dependent and -independent
GMM
UBM systems, with the latter giving better performance. This may be due to the smaller amount of training
data available for gender-dependent age modelling, or because for very young children gender does not provide sig-
nificant benefit for age-group-ID. Subsequent experiments use gender-independent modelling.
TagedPThe next two rows of Table 10 present age-group-ID accuracy obtained with gender-independent GMM
SVM
and i-vector systems. The number of GMM components in the GMM
UBM, GMM
SVM, and i-vector systems is
1024, 512 and 256, respectively, and the dimension of the i-vector total variability subspace is 400. These parameters
were set empirically using a separate evaluation set. The best performance, 82.62%, was obtained using the i-vector
system.
TagedPBased on the results in Section 4, we also conducted age-group-ID experiments using speech that was band-lim-
ited to 5.5 kHz. For the gender-independent i-vector systems, this resulted in an age-group-ID accuracy of 85.77%,
compared with 82.62% for full-bandwidth speech.
TagedPTable 11 is the confusion matrix for the i-vector system with 5.5 kHz bandwidth speech. Each row corresponds to
a grade and shows the percentages of children in that grade who were classified as being in AG1, AG2, and AG3.
152 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

Table 10
Age-group-ID recognition rate (in %) obtained by
the gender-independent GMM
UBM,
GMM
SVM and i-vector systems and gender-
dependent GMM
UBM system.
System Age-group-ID rate (%)

GMM
UBM (gender dep.) 71.76

GMM
UBM (gender indep.) 82.01
GMM
SVM (gender indep.) 79.77
i-vector (gender indep.) 82.62

Table 11
Confusion matrix for age-group identification (in %) for three age-
groups, obtained with the i-vector system using 5.5 kHz band-lim-
ited speech.
Age- Male Female
Group Grade AG1 AG2 AG3 AG1 AG2 AG3

K 100 0 0 100 0 0
1st 100 0 0 100 0 0
AG1 2nd 97.43 2.56 0 97.87 2.12 0
3rd 85.71 10.20 4.08 92.10 7.89 0

4th 33.33 60.60 6.06 38.70 61.29 0

5th 8.57 82.85 8.57 11.42 82.85 5.71
AG2 6th 6.97 81.39 11.62 10.00 80.00 10.00
7th 0 54.83 45.16 2.70 72.97 24.32

8th 0 0 100 0 29.03 70.96

AG3 9th 2.17 6.52 91.30 5.00 20.00 75.00
10th 0 0 100 0 30.00 70.00

TagedPThe thin horizontal lines indicate the AG1, AG2, and AG3 boundaries. The table shows similar characteristics for
boys and girls up to the 7th grade, with the majority of errors near age-group boundaries. At the boundary between
AG1 and AG2, 10% of 3rd grade boys (AG1) are incorrectly classified as AG2, and 33% of 4th grade boys (AG2)
are incorrectly classified as AG1. For girls, the corresponding figures are 8% and 39%. For 7th grade, (AG2) 45% of
boys and 24% of girls are classified as being in AG3, while for 8th grade, (AG3) 29% of girls are classified as AG2,
but none of the boys are misclassified. The inconsistency between the results for boys and girls at the AG2 and AG3
boundary may be because AG3 contains speech from a number of boys whose voices have broken. It may be that
gender-dependent modelling is needed for AG3, even though it is not advantageous overall, or that, as in the case of
gender-ID it is necessary to build separate models for AG3 boys whose voices have or have not broken.

TagedP5.3.2. Human performance

TagedPThe age-group-ID rates for each age-group for human listeners are presented in Table 12. The results indicate that
most confusions arise for test utterances in AG2. Only 50.8% of these are correctly identified. The confusion

Table 12
Confusion matrix for age-group-
ID (in %) for three age-groups,
obtained by human listeners.
Model

Test data AG1 AG2 AG3

AG1 81.2 16.9 1.8

AG2 25.5 50.8 23.6
AG3 3.8 24.4 71.7
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 153

TagedPbetween AG1 and AG3 is small, with only 1.8% and 3.8% of the test utterances from AG1 and AG3 being incor-
rectly identified as AG3 and AG1, respectively. The average performance over all age-groups is 67.54%.

6. Conclusions

TagedPThe objectives of this paper are to investigate, and provide performance benchmarks for, automatic speaker, gen-
der and age-group identification for children’s speech, and to identify the regions of the spectrum that contain the
most important information for these three classification problems.
TagedPThe OGI “Kids” corpus was used in all of the experiments. In order to study the effects of age, it was partitioned
into three age-groups, AG1 (5
8 years), AG2 (8
12 years) and AG3 (12
16 years).
TagedPThree different types of classifier were considered, namely GMM
UBM, GMM
SVM and i-vectors with
PLDA-based classification. Gender and age-group identification rates for human listeners are also presented.
TagedPFor speaker identification and verification for children, the most useful frequencies are up to 1.1 kHz (B1) and
between 2.1 kHz and 5.5 kHz (B3). These are similar to the bands identified in Safavi et al. (2012) and elsewhere for
adults, increased by approximately 38% (B1) and 11% (B3) for children. For gender identification the most useful
frequencies are between 0.9 kHz and 2.6 kHz. This corresponds approximately to the location of the second formant
for vowels, which again is consistent with results for adults’ speech (Wu and Childers, 1991). The most useful fre-
quencies for age-group identification from child speech are between 1.6 kHz and 3 kHz.
TagedPThe results for gender identification show very poor performance for sub-bands 1
11 for the youngest children
(AG1). Since the relationship between the mel and Hertz scales is approximately linear over this region, the equiva-
lent rectangular bandwiths of the triangular mel scale filters is small compared with the wide spacing of spectral har-
monics for these children, due to high fundamental frequency. This increases variability in the high-order MFCCs
(Ghai, 2011). The problem was alleviated by setting the minimum bandwidth of the mel scale filters to be 300 Hz.
The same problem does not seem to occur for speaker or age-group identification, presumably because in these cases
fundamental frequency is relevant for distinguishing between classes.
TagedPFor full-bandwidth speech the best speaker recognition performance is obtained with i-vectors combined with a
PLDA classifier. This achieves verification equal error rates (identification accuracies) of 1.59%, 0.72% and 0.53%
(78.16%, 91.13% and 95.38%) for AG1, AG2 and AG3, respectively.
TagedPFor gender identification the situation is more complex. The best full-bandwidth gender identification accuracy is
79.18% obtained with an age-independent GMM
SVM system. However, the relationship between accuracy and
age is not straightforward. For a GMM
UBM system, accuracy increases from 76.8% for AG1 to 84.1% for AG2,
but then drops to 75.9% for AG3. This is because AG3 includes boys whose voices have and have not broken. When
the training set for boys was partitioned into two sets, corresponding to broken and unbroken voices, and two sepa-
rate boys’ gender models were created, gender identification accuracy for AG3 rose to 87.4%. The average gender
identification performance for human listeners over the three age-groups is 67%, which is poorer than the best auto-
matic system.
TagedPFor age-group identification with full-bandwidth speech, the best performance, 83%, is obtained with a gender-
independent i-vector system. Reducing the speech bandwidth to 5.5 kHz (motivated by the sub-band age-group iden-
tification results) increases the accuracy to 85.8%. The corresponding figure for human listeners is 67.5%.
TagedPIn summary, our results emphasise the importance of taking properties of children’s speech into account in the
design of automatic speaker-, gender- and age-group-identification systems. These include the occurrence of struc-
ture at higher frequencies, the effects of higher fundamental frequency on feature extraction, and the effects of the
onset puberty.
TagedPFinally, speaker and language recognition, in particular, are the subjects of considerable ongoing research. It will
be interesting to discover how well new techniques that emerge are able to accommodate the particular challenges
of children’s speech.

References

TagedPNIST, 2010. 2003 NIST Speaker Recognition Evaluation LDC2010s0. National Institute of Standards and Technology, (Speech Group),
Gaithersburg, MD.
154 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

TagedPAuckenthaler, R., Carey, M.J., Lloyd-Thomas, H., 2000. Score normalisation for text-independent speaker verification systems. Digit. Signal Pro-
cess. 10 (1
3), 42–54.
TagedPBahari, M., Van Hamme, H., 2011. Speaker age estimation and gender detection based on supervised non-negative matrix factorization. In: Pro-
ceedings of the IEEE Workshop on Biometric Measurements and Systems for Security and Medical Applications, pp. 1–6.
TagedPBahari, M.H., McLaren, M., van Hamme, H., van Leeuwen, D.A., 2012. Age estimation from telephone speech using i-vectors. In: Proceedings of
the Interspeech Annual Conference of the International Speech Communication Association, ISCA. Portland, Oregon, USA.
TagedPBatliner, A., Hacker, C., Steidl, S., Noth, E., D’Arcy, S., Russell, M., Wong, M., 2004. ‘you stupid tin box’
children interacting with the aibo
robot: a cross-linguistic emotional speech corpus. In: Proceedings of the 4th International Conference on Language Resources and Evaluation,
LREC. Lisbon, Portugal, pp. 171–174.
TagedPBesacier, L., Bonastre, J.F., Fredouille, C., 2000. Localization and selection of speaker-specific information with statistical modeling. Speech
Commun. 31 (2
3), 89–106.
TagedPBiadsy, F., Hirschberg, J., Ellis, D.P.W., 2011. Dialect and accent recognition using phonetic-segmentation supervectors. In: Proceedings of the
Interspeech Annual Conference of the International Speech Communication Association. Florence, Italy, pp. 745–748.
TagedPBocklet, T., Maier, A., Bauer, J.G., Burkhardt, F., Noth, E., 2008. Age and gender recognition for telephone applications based on GMM supervec-
tors and support vector machines. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-
ICASSP. Las Vegas, Nevada, USA, pp. 1605–1608.
TagedPCampbell, J., Shen, W., Campbell, W., Schwartz, R., Bonastre, J.-F., Matrouf, D., 2009. Forensic speaker recognition. IEEE Signal Process. Mag.
26 (2), 95–103.
TagedPCampbell, W., Karam, Z.N., 2009. A framework for discriminative SVM/GMM systems for language recognition. In: Proceedings of the Inter-
speech Annual Conference of the International Speech Communication Association. Brighton, UK.
TagedPCampbell, W., Sturim, D., Reynolds, D., Solomonoff, A., 2006. SVM based speaker verification using a GMM supervector kernel and NAP vari-
ability compensation. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-ICASSP,
Vol. 1. Toulouse, France, p. I.
TagedPChen, C.-C., Lu, P.-T., Hsia, M.-L., Ke, J.-Y., Chen, O.-C., 2011. Gender-to-age hierarchical recognition for speech. In: Proceedings of the IEEE
International Midwest Symposium on Circuits and Systems, pp. 1–4. August
TagedPDavis, S.B., Mermelstein, P., 1980. Comparison of parametric representations for monosyllabic word recognition in continuously spoken senten-
ces. IEEE Trans. Acoust. Speech Signal Process. 28 (4), 357–366.
TagedPDehak, N., Kenny, P., Dehak, R., Dumouchel, P., Ouellet, P., 2011a. Front-end factor analysis for speaker verification. IEEE Trans. Audio Speech
Lang. Process. 19 (4), 788–798.
TagedPDehak, N., Kenny, P., Dehak, R., Dumouchel, P., Ouellet, P., 2011b. Front-end factor analysis for speaker verification. IEEE Trans. Audio Speech
Lang. Process. 19 (4), 788–798.
TagedPDeMarco, A., Cox, S.J., 2012. Iterative classification of regional British accents in i-vector space. In: Proceedings of the Symposium on Machine
Learning in Speech and Language Processing.
TagedPDobry, G., Hecht, R.M., Avigal, M., Zigel, Y., 2009. Dimension reduction approaches for SVM based speaker age estimation. In: Proceedings of
the Interspeech Annual Conference of the International Speech Communication Association, ISCA. Brighton, UK, pp. 2031–2034.
TagedPDoddington, G., 2001. Speaker recognition based on idiolectal differences between speakers. In: Proceedings of European Conference on Speech
Communication and Technology, Eurospeech ’01. Aalborg, Denmark.
TagedPElenius, D., Blomberg, M., 2005. Adaptation and normalization experiments in speech recognition for 4 to 8 year old children. In: Proceedings of
European Conference on Speech Communication and Technology, Interspeech. Lisbon, Portugal, pp. 2749–2752.
TagedPFringi, E., Lehman, J., Russell, M., 2015. Evidence of phonological processes in automatic recognition of children’s speech. In: Proceedings of
Eurospeech Conference on Speech Communication and Technology, Interspeech. Dresden, Germany, pp. 1621–1624.
TagedPGarcia-Romero, D., Espy-Wilson, C., 2011. Analysis of i-vector length normalization in speaker recognition systems. In: Proceedings of the Con-
ference of the International Speech Communication Association. Florence, Italy, pp. 249–252.
TagedPGauvain, J.-L., Lee, C., 1994. Maximum a-posteriori estimation for multivariate Gaussian mixture observations of Markov chains. IEEE Trans.
Speech Audio Process. 2, 291–298.
TagedPGerosa, M., Giuliani, D., Brugnara, F., 2005. Speaker adaptive acoustic modeling with mixture of adult and children speech. In: Proceedings of the
Annual Conference of the International Speech Communication Association, Interspeech. Lisbon, Portugal, pp. 2193–2196.
TagedPGerosa, M., Giuliani, D., Brugnara, F., 2007. Acoustic variability and automatic recognition of children’s speech. Speech Commun. 49, 847–860.
TagedPGerosa, M., Lee, S., Giuliani, D., Narayanan, S., 2006. Analysing children’s speech: an acoustic study of consonants and consonant-vowel transi-
tion. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-ICASSP, Vol. 1. Toulouse,
France, pp. 393–396.
TagedPGhai, S., 2011. Addressing Pitch Missmatch for Children’s Automatic Speech Recognition Ph.d. thesis. Indian Institute of Technology Guwahati.
TagedPHanani, A., Russell, M., Carey, M., 2013. Human and computer recognition of regional accents and ethnic groups from British English speech.
Comput. Speech Lang. 27 (1), 59–74.
TagedPKenny, P., Boulianne, G., Ouellet, P., Dumouchel, P., 2007. Joint factor analysis versus eigenchannels in speaker recognition. IEEE Trans. Audio
Speech Lang. Process. 15 (4), 1435–1447.
TagedPKinnunen, T., Li, H., 2010. An overview of text-independent speaker recognition: from features to supervectors. Speech Commun. 52, 12–40.
TagedPLee, S., Potamianos, A., Narayanan, S., 1997. Analysis of children’s speech: duration, pitch and formants. In: Proceedings of the Conference on
Speech Communication and Technology, Eurospeech. Rhodes, Greece.
TagedPLee, S., Potamianos, A., Narayanan, S., 1999. Acoustics of children’s speech: developmental changes of temporal and spectral parameters. J.
Acoust. Soc. Am. 105 (3), 1455–1468.
 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156 155

TagedPLee, S., Rose, R., 1998. A frequency warping approach to speaker normalization. In: Proceedings of the IEEE International Conference on Acous-
tics, Speech and Signal Processing, IEEE-ICASSP, Vol. 6. Seattle, WA, pp. 49–60.
TagedPLei, Y., Scheffer, N., Ferrer, L., McLaren, M., 2014. A novel scheme for speaker recognition using a phonetically-aware deep neural network. In:
Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-ICASSP. Florence, Italy, pp. 1714–
1718.
TagedPLiao, H., Pundak, G., Siohan, O., Carroll, M.K., Coccaro, N., Jiang, Q.-M., Sainath, T.N., Senior, A., Beaufays, F., Bacchiani, M., 2015. Large
vocabulary automatic speech recognition for children. In: Proceedings of the Annual Conference of the International Speech Communication
Association, Interspeech. Dresden, Germany.
TagedPLopez-Moreno, I., Gonzalez-Dominguez, J., Pichot, O., Martinez, D., Gonzalez-Rodriguez, J., Moreno, P., 2014. Automatic language identifica-
tion using deep neural networks. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-
ICASSP. Florence, Italy, pp. 5374–5378.
TagedPMahmoodi, D., Soleimani, A., Marvi, H., Razzazi, F., Taghizadeh, M., Mahmoodi, M., 2011. Age estimation based on speech features and support
vector machine. In: Proceedings of the Conference on Computer Science and Electronic Engineering, pp. 60–64.
TagedPMassaro, D.W., 1998. Perceiving Talking Faces: From Speech Perception to a Behavioral Principle. MIT Press, Cambridge, MA.
TagedPMetze, F., Ajmera, J., Englert, R., Bub, U., Burkhardt, F., Stegmann, J., Muller, C., Huber, R., Andrassy, B., Bauer, J., Littel, B., 2007. Compari-
son of four approaches to age and gender recognition for telephone applications. In: Proceedings of the IEEE International Conference on
Acoustics, Speech and Signal Processing, IEEE-ICASSP, Vol. 4. Honolulu, Hawai.
TagedPNaik, J., Netsch, L., Doddington, G., 1989. Speaker verification over long distance telephone lines. In: Proceedings of the IEEE International Con-
ference on Acoustics, Speech and Signal Processing, IEEE-ICASSP, Vol. 1. Glasgow, Scotland, pp. 524–527.
TagedPOrman, O., Arslan, L., 2001. Frequency analysis of speaker identification. In: Proceedings of the Speaker Recognition Workshop: A Speaker
Odyssey. Crete, Greece.
TagedPParris, E.S., Carey, M.J., 1994. Discriminative phonemes for speaker identification. In: Proceedings of the International Conference on Spoken
Language Processing. Yokohama, Japan.
TagedPPelecanos, J., Sridharan, S., 2001. Feature warping for robust speaker verification. In: Proceedings of the Speaker Recognition Workshop: A
Speaker Odyssey. Crete, Greece.
TagedPPrince, S., Elder, J., 2007. Probabilistic linear discriminant analysis for inferences about identity. In: Proceedings of the IEEE International Con-
ference on Computer Vision, pp. 1–8.
TagedPReynolds, D., 1995. Speaker identification and verification using Gaussian mixture speaker models. Speech Commun. 17, 91–108.
TagedPReynolds, D., Rose, R., 1995. Robust text-independent speaker identification using Gaussian mixture speaker models. IEEE Trans. Speech Audio
Process. 3, 72–83.
TagedPReynolds, D.A., Quatieri, T.F., Dunn, R.B., 2000. Speaker verification using adapted Gaussian mixture models. Digit. Signal Process. 10 (1
3),
19–41.
TagedPRogol, A.D., Clark, P.A., Roemmich, J.N., 2000. Growth and pubertal development in children and adolescents: effects of diet and physical activ-
ity. Am. J. Clin. Nutr. 72 (2), 521s–528s.
TagedPRussell, M., Series, R., Wallace, J., Brown, C., Skilling, A., 2000. The star system: an interactive pronunciation tutor for young children. Comput.
Speech Lang. 14 (2), 161–175.
TagedPRussell, M.J., D’Arcy, S., Qun, L., 2007. The effects of bandwidth reduction on human and computer recognition of children’s speech. IEEE Sig-
nal Process. Lett. 14 (12), 1044–1046.
TagedPSafavi, S., Hanani, A., Russell, M.J., Jan
covi
c, P., Carey, M., 2012. Contrasting the effects of different frequency bands on speaker and accent
identification. IEEE Signal Process. Lett. 19 (12), 829–832.
TagedPSarkar, A.K., Do, C.-T., Le, V.-B., Barras, C., 2014. Combination of cepstral and phonetically discriminative features for speaker verification.
IEEE Signal Process. Lett. 21 (9), 1040–1044.
TagedPSchuller, B., Batliner, A., 2013. Computational paralinguistics: emotion. Affect and Personality in Speech and Language Processing. Wiley.
TagedPSchuller, B., Batliner, A., Steidl, S., Seppi, D., 2011. Recognising realistic emotions and affect in speech: state of the art and lessons learnt from
the first challenge. Speech Commun. 53 (9
10), 1062–1087.
TagedPSchuller, B., Steidl, S., Batliner, A., Burkhardt, F., Devillers, L., M€uller, C., Narayanan, S., 2013. Paralinguistics in speech and language
state-
of-the-art and the challenge. Comput. Speech Lang. Spec. Issue Paralinguist. Nat. Speech Lang. 27, 4–39.
TagedPSerizel, R., Giuliani, D., 2017. Deep-neural network approaches for speech recognition with heterogeneous groups of speakers including children.
Natural Language Engineering 23 (3), 325–350.
TagedPShobaki, K., Hosom, J.P., Cole, R.A., 2000. The OGI kids’ speech corpus and recognizers. In: Proceedings of the International Conference on
Spoken Language Processing. Beijing, China.
TagedPSinger, E., Torres-Carrasquillo, P.A., Reynolds, D.A., McCree, A., Richardson, F., amd Dehak, N., Sturim, D.E., 2012. The MITLL NIST LRE
2011. In: Proceedings of the Workshop on Speaker and Language Recognition, Odyssey’12. ISCA, Singapore, pp. 209–2015.
TagedPSolomonoff, A., Quillen, C., Campbell, W., 2004. Channel compensation for SVM speaker recognition. In: Proceedings of the Workshop on
Speaker and Language Recognition, Odyssey’04. Toledo, Spain, pp. 219–226.
TagedPSong, Y., Jiang, B., Bao, Y., Wei, S., Dai, L.-R., 2013. I-vector representation based on bottleneck features for language identification. Electronics
Letters21st 49 (24), 1569–1570.
TagedPVair, C., Colibro, D., Castaldo, F., Dalmasso, E., Laface, P., 2006. Channel factors compensation in model and feature domain for speaker recogni-
tion. In: Proceedings of the Workshop on Speaker and Language Recognition, Odyssey’06. San Juan, Puerto Rico, pp. 1–6.
TagedPVan Heerden, C., Barnard, E., Davel, M., van der Walt, C., van Dyk, E., Feld, M., Muller, C., 2010. Combining regression and classification meth-
ods for improving automatic speaker age recognition. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal
Processing, IEEE-ICASSP. Dallas, TX, USA, pp. 5174–5177.
156 S. Safavi et al. / Computer Speech & Language 50 (2018) 141
156

TagedPVariani, E., Lei, X., McDermott, E., Lopez-Moreno, I., GonzalezDominguez, J., 2014. Deep neural networks for small footprint text-dependent
speaker verification. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE-ICASSP. Flor-
ence, Italy, pp. 4080–4084.
TagedPWan, V., Campbell, W., 2000. Support vector machines for speaker verifcation and identifcation. In: Proceedings of the IEEE Signal Processing
Society Workshop on Neural Networks for Signal Processing, Vol. 2, pp. 775–784.
TagedPWilpon, J., Jacobsen, C., 1996. A study of speech recognition for children and the elderly. In: Proceedings of the IEEE International Conference on
Acoustics, Speech and Signal Processing, IEEE-ICASSP. Atlanta, GA.
TagedPWong, E., Sridharan, S., 2001. Fusion of output scores on language identification system. In: Proceedings of the Workshop on Multilingual Speech
and Language, Processing. Aalborg, Denmark.
TagedPWu, K., Childers, D.G., 1991. Gender recognition from speech. Part II: fine analysis. J. Acoust. Soc. Am. 90 (4), 1841–1856.
TagedPYamada, T., Wang, L., Kai, A., 2013. Improvement of distant-talking speaker identification using bottleneck features of DNN. In: Proceedings of
the Annual Conference of the International Speech Communication Association, Interspeech. Lyon, France, pp. 3661–3664.
TagedPZissman, M., 1996. Comparison of four approaches to automatic language identification of telephone speech. IEEE Trans. Speech Audio Process.
4 (1), 31–44.

View publication stats