Docstoc

SUBJECTIVE AGE ESTIMATION OF TEL

Document Sample
SUBJECTIVE AGE ESTIMATION OF TEL Powered By Docstoc
					SUBJECTIVE AGE ESTIMATION OF TELEPHONIC VOICES

Loredana Cerrato, Mauro Falcone, Andrea Paoloni

Fondazione Ugo Bordoni
via Baldassarre Castiglione 59, 00149 Roma

Tel: +39.06.54801
Fax : +39.06.5480.4405
Email: {loredana, falcone, pao}@fub.it



Abstract
The aim of this study was to investigate the extent to which listeners can judge some unknown
speakers' characteristics. In particular we are interested in investigating the speaker accuracy of
assessing age and gender by only hearing the voice recorded over the telephone line. Moreover we
tried to evaluate the reliability of subjective age estimation in the forensic context.
The results of the statistical analysis we carried out show that listeners are capable of assigning a
general chronological age category to a voice without seeing or knowing the talker and they are
able to distinguish between male and female voices transmitted over the telephone line.


Introduction
Everyday experience allows us to state that it is possible to make an estimate of some of the
characteristics of an unknown speaker by only listening to his/her voice. Some of these
characteristics are related to speakers’ physiological characteristics, such as gender, age, weight,
some other depend on their cultural and social background: geographical provenience, level of
education, and also emotional state.
Many studies have examined the extent to which listeners can judge a speaker's characteristics from
voice cues only. The results show that gender can be accurately judged (Mullenix, 1995), for height
and weight assessment there are different results: Lass (Lass et alii, 1980) reported that these two
speakers' characteristics could be easily assessed even in different conditions of signal distortion;
while more recently Kunzel (Kunzel, 1989) reported negative results for the reliability of height
and weight estimation. Listeners appear to be able to judge also if a speaker is sober or intoxicated
(Pisoni, 1989).
Several researchers reported that listeners are able to make fairly accurate judgements about
speakers' age from voice cues only. Both the results of tests requiring the subjects to indicate the
age group of the target voice (Neiman 1990, this study) and to provide a direct age estimate (Ryan
and Burk, 1974; Hory and Ryan, 1981; Braun, 1996), show consistent results indicating that
estimates of age are generally accurate within a decade.
Most of the previous studies on subjective age estimation used speech samples produced under
controlled phonation, recorded in normal conditions (Shipp and Hollien, 1969; Linville and Fisher
1985; Ryan and Burk, 1974; Hory and Ryan, 1981; Neiman and Applegate, 1990; Braun, 1996).




                                                 1
We thought it could be also interesting to use voice samples produced in a natural way and
recorded from telephone calls, as the application we envisage is forensic voice profiling in forensic,
in particular for the critical task of identification of voices recorded by means of phone tapping .

Experimental set up
We used a set of 42 voice samples, which were equally balanced between the genders with 21
male and 21 female voices.
The samples were randomly selected from the SIVA database (Speaker Identification and
Verification Archive) gathered over the Italian public telephone line and consisting of 2000 phone
calls made by 1000 speakers, equally distributed by sex, age and geographic area (Cerrato, 1997).
The 42 voices were almost equally distributed in the three main Italian geographic areas: Northern,
Southern and Central Italy, which coincide with the three main regional varieties of the Italian
language. Italian is characterized by many dialects and regional varieties; the broadest way to point
out the large geographical variability of the Italian language is to identify three main regional
varieties, which correspond to the three main geographical subdivisions of the Italian territory
(Rohlfs, 1996).
Each speaker read the same short passage of the length of about 40 seconds, consisting of a
fictitious curriculum vitae, from which it was not possible to deduce any personal information
related to the speaker.
Normally the response to the subjective age estimation task can be done in two different ways:
indicating the age group of the target voice or providing a direct age estimate.
For our investigation we followed the former procedure, establishing 7 age groups of 7 year-range
as follow: 18-24, 25-31, 32-38, 39-45, 46-52, 53-59, 60-66.
3 male and 3 female speakers were selected for each age group for a total of 42 speakers.
17 listeners (male and female mostly university students) with no hearing pathologies took part to
our experiment; they had no training in perceptual phonetics. They were paid the equivalent of
30Euro to carry out the test of the total length of 40 minutes, which was submitted in two parts with
an interval of 10 minutes.
The subjects listened, over headphones at a comfortable listening level (about 73 dBA at the ear
reference point), to the 42 voices reading the same passage. They were required to indicate the
perceived age group and the gender of each speaker, ticking the appropriate answer among the ones
given in the answer sheet.


Analysis of the data
The data were analysed by means of statistical analysis in order to investigate the magnitude and
the distribution of the errors in relation to the speaker’s age. This has been done by means of
standard analysis and with the support of a series of graph and tables.
A comparison of the obtained results with the achievement of other researchers has also been
performed, in order to check the reliability of our experiment and, more importantly, to allow a
comparison between data gathered from telephonic speech and normal speech.
Several researchers (Neiman 1990, Braun, 1996) believe that a possible representation of the
accuracy of perceived age estimation can be given by correlation or a similar estimator, which are
ad hoc relationships between chronological age (CA) and perceived age (PA). When listeners are
constrained to give their judgement according to age groups, chronological age is represented by
the central value of the age class, while perceived age is represented by the mean value of all the



                                                  2
answers given by all the listeners. The percent of correspondence between CA and PA is always
calculated dividing the lowest of these two values by the highest ones, thus obtaining a result
necessarily minor or equal to 1. For example if the CA of a sample is 18-24 (mean value 21) and
the PA is 32, the percent of correspondence is 21/32, that is 0.65 or as commonly reported a 65% of
accuracy.
As a consequence of this formulation a percent correspondence close to the unit means that
listeners correctly judge the speaker’s age. On the contrary, as the value approaches 0, the
estimation of the speaker's age gets progressively worse.
In table 1 we show the percent correspondence between CA and PA in our experiment. In table 2,
we summarize the values obtained by other researchers (for non telephonic voices).
In the test we run,we find quite high values ranging from 68% and 78% that are the worst cases in
correspondence of youngest and oldest age group, and reaching the highest value of 96% in the
central group.
A more widely spread and common measure, that reflects the extent of a linear relationship
between two data sets, is the dimensionless index that ranges from -1.0 to 1.0 named Pearson
product moment correlation. Also in this case the best result is represented by the value 1. This
measure has been used to calculate the correlation between perceived age and chronological age by
different researchers, on high quality speech material, as reported in table 3. In our case the Pearson
correlation gives a coefficient equal to 0.77. Notwithstanding the same aim of these two measures,
the ranking of subject judgment ability may not, as in the case, be the same as reported in table 2
and 3. For instance, the Pearson correlation in (Neiman, 1996) is higher than Pearson correlation in
(Braun 1996), while for the percent correct correspondence we have an opposite situation.
These measures only evaluate the subject’s judgment trend in relation to the distribution of
speaker’s age and do not take into account the extent of the errors, which we may be interested in.
In fact we analysed more in detail the errors, also in relation to the speaker’s age, and we
investigated the effect of over and under estimation by listeners. In order to do this, we had to
perform further analyses as the previous results were not able to describe that specific aspect.
It is possible to consider 2 errors: the absolute error and the relative error. The former is more
suitable to estimate the extent to which a subject is able to correctly assess a speaker voice, as it
integrates errors as positive values without considering the sign of the distance between estimated
age and chronological age. The latter is more suitable when we are interested in the extent to which
a class of listeners is able to properly assign an age to an unknown voice. In this case it is possible
to balance positive and negative errors, as we are looking at the estimate given by the pool of
subjects rather than the estimate of single listeners. This last measure is more appropriate in the
forensic environment. A simple but effective way to perform this analysis, is the mapping of
perceived age and its error for each age class, as reported in figure1. It is clear from this graph that
listeners tend to overestimate young speakers and to underestimate old speakers. This tendency was
pointed out also by Ship and Hollien (Ship and Hollien, 1969) in their experiment in which subject
performed a direct age estimation task.
As we performed an age estimation test constrained by age groups, we also calculated the
percentage of correct answer for each age class, reported in figure 2.
The phenomenon of overestimation of young speakers and of underestimation of older speakers is
more evident in the detailed set of figures 2a-2g, where we report the answer distribution for each
of the seven age classes. The results show that listeners assigned the correct age class mainly to the
age group 46-52 (figure 2e). Listeners very rarely guessed the age of young speakers (figure 2a)
which were overestimated. Older speakers (figure 2f and2g) were underestimated, in particular for



                                                   3
the age group 60-66 (figure 2g) most of the answers fall in the age group 46-52, that is an
underestimation of 14 years.
This kind of representation (figure 2) seems to outline that the listeners’ performance is not very
good, as on average the overall percentage of correct answers we obtain is only 27%.
It shouldn't be a surprise that these two performed analyses give apparently inconsistent results, in
fact correlation-like analyses give quite good results, while the value of overall percentage of
correct answers is very low. Again we wish to underline that these two analyses aim at different
investigation so their results do not necessarily have to be similar
The results of the statistical analysis we carried out show that listeners were remarkably correct at
estimating the age of some speakers, but not that of others. In general they are able to assign an
unknown voice to an age group, in fact if the age classes had a bigger range (21 years) the
percentage of correct answers would result higher. This is suggested by the representation reported
in figure 3, where both the percentages of correct answers for each age class and the percentage of
answers given to the adjacent age classes are plotted.
Finally if we consider the speakers as forming only three groups: younger speakers, adult speakers
and older speakers, we can see that the overall percent of correct answers rise up to 71%, as shown
in figure 4.



Judgements relative to gender.
The other goal of our experiment was to determine whether listeners are able to distinguish
between male and female voice of unknown speaker while hearing his/her voice over the phone.
Our results confirm the well established ability of listeners to distinguish between male and female
voice. In fact the percentage of correct answer relatively to the gender of the speakers is 100%,
which means that none of the listeners ever doubted of the gender of the speakers heard in the voice
samples.


Discussion
The main aim of this study was the evaluation of the ability of listeners to estimate the age of
unknown speakers in a subjective estimation task and the reliability of this subjective measure for
forensic purposes. The estimation of speakers' age is one of the most frequent tasks in speaker
profiling for forensic purposes: whenever an anonymous recording is submitted for analysis, an
estimation of the speaker's age and sometimes also of some other of his/her characteristics, are
requested by investigators. This usually results in a major lead in the identification and search of
the person (Nolan 1996), for this reason it is important to understand the reliability of this measure.
Our results show that estimates of age are generally accurate only for broad age classes and confirm
the intuitive expectation that people are able to assign a general chronological age category to a
voice without seeing or knowing the talker. Therefore in order to accomplish the task of age
estimation in forensics, it is advisable to use general descriptions such as "young", "adult", "old" or
similar, rather than the indication of age using age groups.
Moreover our result show that listeners are able to correctly recognize the gender of an unknown
speakers.




                                                  4
REFERENCE
Braun, A., 1996. Age estimation by different listener groups, The Journal of Speech Language and
    the Law, vol. 3 n. 1 , pp.65-73
Cerrato, L. 1997: Sulle informazioni contenute nel SIVA, un database per la verifica e
l’identificazione del parlatore, Relazione FUB 5b05197, raccolta pubblicazioni FUB 1997 p.61-70
Horii, Y., Ryan, W.J., 1981. Fundamental Frequency Characteristics and Perceived Age of Adult
    Male Speakers, Folia Phoniatrica, vol. 33, pp. 227-233
Kunzel, H., 1989. How well does average fundamental frequency correlate with speaker height and
    weight?, Phonetica, vol. 46, pp. 117-125
Lass, N.J., Phillips, J.K., Bruchey, C.A., 1980. The effect of filtered speech on speaker height and
    weight identification, Journal of Phonetics, vol. 8, pp. 91-100
Linville, S.E., Fisher, H.B., 1985. Acoustic characteristics of perceived versus actual vocal age in
    controlled phonation by adult females, JASA, vol. 78, pp. 40-48
Mullenix, J.W., Johnson, K.A., Topcu-Durgun, M., Farnsworth, L.M., 1995. The perceptual
    representation of voice gender, JASA, vol. 98, pp. 3080-3095
Neiman, G.S., Applegate, J.A., 1990. Accuracy of listener judgements of perceived age relative to
    chronological age in adults, Folia Phoniatrica, vol. 42, pp. 327-330
Nolan, F., 1996. Speaker recognition and forensic phonetics, in "The handbook of Phonetic
    sciences" by Hardcastle W., Laver J., pp. 744-766
Pisoni, D.B., Martin, C.S., 1989. Effects of alcohol on the acoustic phonetic properties of speech:
    perceptual and acoustic analyses, Alcoholism clinical and experimental research, vol. 13, pp.
    577-587
Rohlfs, G., 1996. Studi e Ricerche su lingua e dialetti d' Italia, Sansoni Editore
Ryan, W.J., Burk, K.W., 1974. Perceptual and Acoustic Correlates of Aging in the Speech of
    Males, Journal of Communication Disorder, vol. 7, pp.181-192
Shipp, T., Hollien, H., 1969. Perception of the Aging Male Voice, JSHR, vol. 12, pp. 703-710




                                                 5
Biographical notes of the authors:
Loredana Cerrato received an honor degree in Linguistics by the University of Naples "Federico II"
in 1992. In 1995 she was visiting scientist at the National Acoustical Laboratories (AHS) in
Chatswood, Sydney, Australia. From August 1996 to April 1999 she has worked at Fondazione
Ugo Bordoni in the Multimedia Communication Division of the Voice Communication Group.
In March 1999 she started a Doctorate course in “Language theories and Speech technology” at the
University of Bari. Her interests cover the fields of speech analysis, speech perception,
sociolinguistics, experimental phonetics, forensic phonetics, speech recognition and automatic
translation. The results of her research lead to the publication of 25 papers in conference
proceedings and scientific journals.


Mauro Falcone received the degree in Physics in 1988. He is a researcher in the Multimedia
Communication Division, in the Speech Communication group at Fondazione Ugo Bordoni. His
interests cover: speech quality evaluation; speech output system assessment; speaker
verification/identification assessment; interactive and speech driven system assessment. He has
published more than 50 works among project reports, articles in books, scientific journal,
conference proceedings.

Andrea Paoloni received the degree in Electronic Engineering from the University of Rome, in
1973. Since 1974 he has been working as researcher at the Fondazione Ugo Bordoni, where he is
now Head of the Speech Communication Group. His research activity covers the fields of speech
processing, with particular emphasis in speaker recognition and systems evaluation. He has
authored 120 papers. He has been working also for the criminal justice system as an expert in
speaker identification and authentication of recordings. He is now also chairman of the COST 250
project (Speaker recognition in telephony).


Table 1 Percent of correspondence between chronological
and perceived age
Age classes            CA      PA      CA vs. PA
18-24                  21      31      68%
25-31                  28      35      80%
32-38                  35      38      92%
39-45                  42      46      91%
46-52                  49      47      96%
53-59                  56      48      86%
60-66                  63      49      78%
                                       84% (mean)




                                                          6
Table 2 Summary of percent correspondence between
chronological and perceived age in previous studies
(*Braun used only male voices)
Authors                    year            result
Neiman, Applegate          1990            80.2%
Braun*-(expert)            1996            85.5%
Braun* (non expert)        1996            84.1%
this study                 1998            84.4%




Table 3 Summary of Pearson correlation coefficient
established in previous studies
Authors                      year        Pearson
                                         coefficient
Ship, Hollien                1969        0.88
Ryan, Burk                   1974        0.77
Horii, Ryan                  1981        0.76
Neiman, Applegate            1990        0.88
Braun (expert)               1996        0.70
Braun (non expert)           1996        0.68
this study                   1998        0.77




   80

   70

   60

   50

   40

   30

   20

   10

    0
        18-24 25-31 32-38 39-45 46-52 53-59 60-66


Fig. 1 Perceived age and its error. Triangles stand for
mean of perceived age, circles represent the central
value of the age class.



                                                          7
   50


   40


   30


   20


   10


    0
        18-24   25-31   32-38   39-45    46-52    53-59    60-66


Fig. 2 Percentage of correct estimates for age classes




   50                                                                 50
   40                                                                 40
   30                                                                 30
   20                                                                 20
   10                                                                 10
    0                                                                  0
         21     28      35      42      49       56   63                   21   28   35   42     49   56   63


Figure 2a – Age class 18-24                                        Figure 2b – Age class 25-31

   50                                                                 50
   40                                                                 40
   30                                                                 30
   20                                                                 20
   10                                                                 10
    0                                                                  0
        21      28      35      42      49       56   63                   21   28   35   42     49   56   63


Figure 2c – Age class 32-38                                        Figure 2d – Age class 39-45


                                                                      8
   50                                                                 50
   40                                                                 40
   30                                                                 30
   20                                                                 20
   10                                                                 10
    0                                                                  0
        21      28      35      42      49       56   63                   21    28   35    42    49   56    63


Figure 2e – Age class 46-52                                        Figure 2f – Age class 53-59

   50
                                                                   Fig.2a-g In these figures the distribution of answers
   40
                                                                   for each of the seven classes is reported. On the
   30                                                              abscissa there is the central value of the age class,
   20                                                              on the ordinate the percentage of answers given by
   10                                                              listeners. The dot on the top of each figure indicates
    0                                                              the correct chronological age.
        21      28      35      42      49       56   63


Figure 2g – Age class 60-66




   90
   80
   70
   60
   50
   40
   30
   20
   10
    0
        18-24   25-31   32-38   39-45    46-52    53-59    60-66


Fig. 3 Percentage of correct estimation (black) and
percentage of answers given to adjacent classes (gray).




                                                                       9
   100

    80

    60

    40

    20

     0

           18-38         32-52         45-66
          "young"       "adult"       "older"

Fig. 4 Percentage of correct answers for broad age classes.




                                                              10

				
DOCUMENT INFO