It is widely acknowledged that distinguishing twins poses a major challenge in the field of forensics because these individuals are physically very similar. For instance, biometrics such as fingerprints (Jain, Prabhakar & Pankanti, 2002) or palmprints (Kong, Zhang & Lu, 2006) have often been investigated in twins to study the subtle differences frequently observed between them. Similarly, researchers have investigated behavioral characteristics of twins such as handwriting (Srihari, Huang, & Srinivasan, 2008). In the same way that handwriting depends on physiology as much as on behavioral factors like training and habits, the foundations of speaker recognition are largely grounded on the idea that a voice is determined not only by anatomical structure but also by nonbiological or behavioral factors. These factors include mainly social or dialectal aspects but other environmental influences are possible. Nolan and Oh (1996, p. 39) highlighted that aspects of personal voice quality are determined by anatomical inheritance, mimicking traits from other people, or else, they are arbitrarily chosen in order to mark someone’s personality. This organic-learned dichotomy (Nolan, 1997; Nolan & Oh, 1996) may be a good translation in phonetic terms of the well-known nature–nurture dichotomy, first outlined by Sir Francis Galton in 1875 (Galton 1875, in Segal 1993, p. 45).


This distinction, nature vs. nurture, has resulted in fruitful twin research in many disciplines, where heritability or concordance rates are calculated for certain traits in order to determine whether these could be genetically influenced. This happens when there is greater similarity on that trait between monozygotic (MZ) twin pairs than between dizygotic (DZ) twins. MZ twin pairs share 100% of their alleles and DZ twins, on average, share only half their genetic information, whereas both types of twin pairs share essentially the same prenatal and postnatal environments (Stromswold 2006, p. 334). This is the essence of the classical twin design, which requires that an important assumption be made: the equal environment assumption (EEA), i.e., it is assumed that the two twin types have similar environmental experience.1 A number of studies have investigated the differences in MZ and DZ twins to assess the effect of genetic factors in voice (see Section 1.2), but—to the best of our knowledge—the joint consideration of MZ, DZ and non-twin siblings2 has not been approached in phonetic studies before.


Acknowledging the existence of these two “forces”, i.e., nature and nurture (alternatively also referred to as organic and learned factors, respectively) to explain the (dis)similarities between twins does not mean that their relative influence or importance can be clearly separated. Moreover, there is a third element, epigenetics, which is often neglected in twins’ studies even though it usually comes into play to explain how changes in gene expression caused by mechanisms other than changes in the underlying DNA sequence can cause divergence in twins, which may account for strikingly dissimilarities between MZ twins. See, for instance, how a particular epigenetic process called DNA methylation (Martino et al. 2013; Philips, 2008) is reported to make the expression of genes weaker or stronger.


A recent study (Felson, 2014) has aimed at undertaking a comprehensive evaluation of the EEA, which has often caused some skepticism amongst researchers. Felson presents evidence that suggests that neither extreme of the opposing views is correct, and that the truth lies somewhere in the middle. In other words, it seems that although environmental similarity may not have been adequately measured in some sociology-related twin studies, “the resulting bias is likely modest” (p. 184). Therefore it could be argued that despite its limitations twin research is still greatly encouraged nowadays to shed light on the interplay of genetic and environmental factors. Particularly referring to the difficult task of searching for genetic influences of the voice, Sataloff (1995) pointed out that “the complexities of genetic research in humans have left most of the relevant questions unanswered” (p. 17).


Studies on twins’ voices are undertaken for at least two main reasons. On the one hand, this type of studies can reveal—for the investigated voice characteristics—how the results of pairwise comparisons vary depending of the type of speaker considered. The comparison is usually between MZ twins and DZ twins; San Segundo (2014) also proposed drawing comparisons against non-twin siblings and unrelated speakers. The genetic influence of the analyzed voice characteristics is apparent when higher similarity is observed in MZ twins than in DZ twins, non-twin siblings or unrelated speakers. On the other hand, the relevance of twin studies to Forensic Phonetics3 in particular lies in the search for robust4 voice characteristics that could facilitate the discrimination of very similar speakers.5 Hence, these four speaker groups (MZ twins, DZ twins, non-twin siblings and unrelated speakers) are proposed for testing the performance of a speaker-comparison system. As can be observed, the two highlighted aspects are strongly linked, since a set of characteristics may be robust for speaker comparison as far as they are maximally influenced by the speaker’s genetic endowment and minimally due to learned factors, the latter favoring voice disguise or imitation. The predominance of genes over environment is clearly related to the two most repeated (and probably important) criteria in the identification of characteristics for Forensic Speaker Comparison (FSC), namely that these characteristics should be as consistent as possible for each speaker (low within-speaker variability) and that they should exhibit large variation amongst speakers (high between-speaker variability). Among others, these criteria were already outlined by Wolf (1972) and Nolan (1983) in the phonetic realm, but they also appear in the literature specifically related to automatic speaker recognition (ASR). For example, Kinnunen and Li (2010) refer to the same characteristics for an ideal ASR system.


From a literature review of around 30 voice-related twin studies (San Segundo, 2014), we can draw some interesting conclusions. For instance, it seems that previous phonetic studies focusing on twins have aimed at basically one of the following objectives (see San Segundo, 2015): (a) trying to find a genetic component in the variation of certain voice characteristics by searching differences between MZ and DZ twin pairs (e.g., Debruyne, Decoster, Van Gijsel, & Vercammen, 2002; Przybyla, Horii, & Crawford, 1992) or else, in a forensic scenario, (b) creating a system capable of discriminating between MZ and DZ twins (e.g., Forrai & Gordos, 1983) or, more frequently, testing whether it is possible to distinguish a speaker from his/her co-twin (e.g., Ariyaeeinia, Morrison, Malegaonkar, & Black, 2008; Homayounpour & Chollet, 1995; Künzel, 2010; Loakes, 2006; Nolan & Oh, 1996; Scheffer, Bonastre, Ghio, & Teston 2004). For a thorough discussion of the results derived from previous twin studies, see San Segundo (2014), where previous works have been classified in four groups depending on whether they represent perceptual, acoustic, articulatory or automatic (ASR) approaches.


While most of the studies undertaken from an acoustic perspective focus on traditional phonetic characteristics, as described in Künzel (2011) and Rose (2006)—for example, fundamental frequency (f ₀), formant patterns or temporal characteristics such as word duration, vowel duration or Voice Onset Time (VOT)—, research into laryngeal features and phonation characteristics derived from the glottal waveform has been very limited. Classical distortion characteristics such as jitter and shimmer have only occasionally been explored in twins (van Lierde, Vinck, De Ley, Clement, & Van Cauwenberge 2005; Weirich & Lancia, 2011). More recently, some investigations on twins’ voices (San Segundo, 2012; San Segundo & Gómez-Vilda, 2013; San Segundo 2014; San Segundo & Gómez-Vilda, 2015) have analyzed a considerably larger number of glottal features, on the basis of the voice analysis methodology described in Gómez-Vilda et al. (2007), which relies on the decoupling of the vocal tract from the glottal source estimates.


If we focus on ASR studies in particular, this approach to twins’ voices has not been extensively developed, in comparison with other acoustic studies investigating specific segmental features. The main objectives of the ASR studies reviewed in San Segundo (2014) are one of the following: (a) comparing the performance of ASR systems with the ability of familiar and non-familiar listeners to discriminate twins (Homayounpour & Chollet, 1995); (b) testing if an ASR system is able to detect correctly the twin pair of a speaker (Scheffer et al. 2004), or (c) in general, testing the intra-speaker, inter-speaker and intra-pair similarity of twins, for example in terms of Likelihood Ratios (LRs) or similarity coefficients. In this last research line we find two recent studies, namely Kim (2010) and Künzel (2010). Since both use the same ASR system that we are using in our study, we will devote an important part of this section to the description of their objectives and main findings.


Kim (2010) studied 22 Korean female twin pairs (17 MZ, including one triplet and five DZ) using Agnitio Voice Biometrics’ Batvox^TM (Version 3.0). Two different speaking styles—text reading and spontaneous interview—were used. The results of this investigation showed that every twin speaker was correctly identified in the same speaking style condition (when models and test files were read speech). According to the author, this would suggest that, at least in ASR, the same speaking style setting should be provided in order to get more confident results. Noteworthy of this study is also that in nine out of 22 pairs, intra-twin LRs in the same speaking style condition were higher than intra-speaker LRs in different speaking style condition. This situation is highly undesirable in a forensic context, where inter-speaker variation should be larger than intra-speaker variation (Wolf, 1972).


Künzel (2010) is the most recent study on automatic speaker recognition in which a Bayes-based system (Batvox^TM, Version 3.1) was used to calculate LR distributions for inter-speaker, intra-pair and intra-speaker comparisons. A total of 35 German MZ pairs (26 female and nine male) participated in this study and two different tests were designed. In the first one, both target voices consisted of the same read text, while in the second one the speaker models were built from spontaneous speech samples but read speech samples were used as targets. The results showed that in the first experiment the automatic system allowed a perfect distinction of each member of a male twin pair (i.e., 0% of Equal Error Rate; EER) and 0.5% EER for female twin pairs. In the second experiment, the EER rose to 11% for male twin pairs and 4.4% for female twin pairs. These values represent the crossover point in the Tippett plot for the inter-speaker/intra-speaker LR distributions. However, the results for female twins are worse when considering intra-pair/intra-speaker distributions (19% EER in the first experiment and 48% EER for the second experiment). Therefore, the performance of the system was clearly superior for male than for female voices. The author’s explanation for this phenomenon is that “as a consequence of the higher fundamental frequency of female voices the spacing of the harmonics is less dense than for male voices, which in turn yields less speech sound- and speaker information in the spectrum” (Künzel, 2010, p. 270). This becomes clearer if we bear in mind that the spectrum is used for the extraction of the mel-frequency cepstrum coefficients (MFCCs), which are the features this automatic system used is based upon.


Finally, references to siblings’ voices within an automatic approach are almost inexistent except for the study of Charlet and Lecha (2007), which tested a text-dependent speaker recognition system with 33 families, finding that the son was highly confused with his brother. The implication is that someone could be a good impostor of his brother, making this type of speakers especially relevant in forensic studies and thus justifying not only the study of twins but also of non-twin siblings.


This investigation is part of a larger research project (San Segundo, 2014); more details about the corpus of twin and non-twin subjects can be found in San Segundo (2013b, 2014). The automatic analysis that we present here is based on speech samples extracted from the fifth corpus task: informal interview with the researcher.


As explained above, we carried out three types of tests, which yielded results for intra-speaker, inter-speaker and intra-pair comparisons. If we first look at the results for intra-speaker and inter-speaker comparisons alone, we see that similarly high coefficients of recognition are obtained for all the pooled four speaker types (MZ, DZ, B and US). This can be observed in Figure 1, which shows a 0% EER. The input values for the creation of this figure were of two types:


The 0% EER indicates that there is no overlap of matches and non-matches, so neither FA nor FR occur. This shows that the overall system performance with high-quality recordings and without taking into account the sibling effect (intra-pair comparisons) is perfect.


When taking into account also intra-pair comparisons, in addition to matches and non-matches, the recognition coefficients are expected to be much lower, as the comparison is not between the same individuals. However, different patterns were observed depending on the type of speaker (MZ, DZ, B or US). This can be seen in Table 1, where the values obtained are classified per speaker (i.e., his intra-speaker coefficients) and per speaker pair (i.e., their intra-pair coefficients), depending on whether they are MZ, DZ, B or US. As it can be observed in this table, all intra-speaker comparisons yield similarly high coefficients of recognition. In relation to the intra-pair comparisons, Table 1 is useful to observe the different values obtained by different speaker pairs, i.e., the performance of the system can be analyzed per speaker or per speaker pair. The fact that the speakers in this investigation are not very numerous is an advantage in order to carry out this kind of detailed examination. For instance, if we look at within-group differences, the value of MZ pair 39–40 (0.64) is very different from the other pairs’ coefficients (much higher in average).


If we are interested in the behavior of the groups in general, and not specifically in each pair, Table 2 and its corresponding figure (Figure 2) are more insightful and probably more appropriate to assess the system performance depending on the speaker type. According to the information in Table 2, MZ intra-pair comparisons yield the highest values (i.e., the dissimilarity is the lowest, so they are the most similar speakers). From the average values obtained by the MZ pairs to the coefficient values yielded for US, we observe a gradation from largest to lowest, all through the average values of the DZ intra-pair comparisons and the B intra-pair comparisons. This trend is thus in agreement with our hypothesis, where we predicted the following scale (from more to less similar): MZ > DZ > B > US. In other words, the coefficient grading goes in the same direction as the “magnitude” of kinship relationship.


We have added to Table 2 the average coefficients obtained in (MZ) intra-speaker comparisons. As expected, these same-speaker comparisons yield the highest coefficients. The inclusion of these matches in the table is intended to serve as a baseline to which the rest of (intra-pair) coefficients can be compared, under the assumption that nobody could be more similar to anyone than to himself, although some exceptions may occur in the case of MZ twins, as we describe in Section 3.3.


The MZ intra-pair comparisons deserve special consideration. As they represent the cases of highest similarity in human beings, they have been more often studied than the other types of kinship relationships considered in this investigation. In the case of FSC carried out using automatic recognition methods, the existence of previous studies that have also used Batvox^TM for the voice comparison of MZ twins gives us the opportunity to compare our results with previous findings.


For the MZ twins participating in our study, we have considered useful to compare the coefficients obtained by each speaker in the intra-speaker (IS) comparisons with the coefficients obtained by these same speakers in the intra-pair (IP) comparisons. Table 3 contains this information, extracted from the general results shown in Table 1.

We have calculated an IS–IP value to measure the difference between the IS comparison coefficient and the IP comparison coefficient. This has been done per speaker and speaker pair. Note however that for the IS coefficients, we have only taken into account the values obtained by one member of the pair: the twin member with the even number in his pair (i.e., 02, 04, 06, 08, etc.). The selection of the IS coefficients of the odd pairs did not yield any negative value. That is the reason why we show the results of the even numbers; as explained above, the interest of this calculation lies in finding any possible speaker pair subject to discrimination errors by the system under test.


As shown in Table 3, only two cases out of 12 MZ pairs show a negative value in their IS–IP value, meaning that the IP coefficient is larger than the IS coefficient. This implies that in these two cases the automatic system Batvox^TM would not be able to discriminate between one twin and the other. In positive values, we can say that in 83.3% of the total MZ cases, the system identifies an identical twin without falsely accepting his co-twin. In Figure 3 we draw the IP and IS coefficient values per MZ twin pair, in IS-decreasing order to show how the trend “large IS–small IP” is followed in all cases except in the last two, corresponding to the MZ pairs 01v02 and 35v36, as we could also observe in Table 3. These two pairs account for the 16.7% not confirming the hypothesis that IS comparisons are always larger than MZ IP comparisons. However, as we will discuss below, this small percentage is in agreement with previous studies.


The two specific cases of MZ twins that were not be recognized by the system explain the 9.9% EER obtained in Figure 4, where the line for matches (right line) is used in this case for intra-pair comparisons (only MZ) and the line for non-matches (left curve) represents the inter-speaker comparisons.


In Figure 5, we have added the curves in Figure 1, which showed the overall system performance. The line further to the right (black) is for IS comparisons of all the speakers in the corpus, and the other right line (blue) represents the IP comparisons, only for MZ. In this new figure, one can distinguish a left-shift from the general IS-curve to the MZ IP-curve, which indicates the performance deterioration from a situation where the system has to recognize same speakers to a situation where identical-twin recognition takes place. The lines for the non-matches in both cases (compare the two curves rising to the left) are practically identical. In both cases, they represent different-speaker comparisons, while in one case (yellow curve, i.e., non-matches in Figure 1) these tests compared the first session of each speaker with the first session of all the other speakers in our corpus; and in the other case (red line, i.e, non-matches in Figure 4), the different-speaker tests were obtained from comparing each speaker’s first session with all the other speakers’ second session.