Machine Assisted Analysis of Vowel Length Contrasts in Wolof
MMachine Assisted Analysis of Vowel Length Contrasts in Wolof
Elodie Gauthier , Laurent Besacier , Sylvie Voisin Laboratoire d’Informatique de Grenoble (LIG), Univ. Grenoble Alpes, Grenoble, France Laboratoire Dynamique Du Langage (DDL), CNRS - Universit´e Aix Marseille, France [email protected], [email protected], [email protected]
Abstract
Growing digital archives and improving algorithms for au-tomatic analysis of text and speech create new research oppor-tunities for fundamental research in phonetics. Such empiricalapproaches allow statistical evaluation of a much larger set ofhypothesis about phonetic variation and its conditioning factors(among them geographical / dialectal variants). This paper il-lustrates this vision and proposes to challenge automatic meth-ods for the analysis of a not easily observable phenomenon:vowel length contrast. We focus on Wolof, an under-resourcedlanguage from Sub-Saharan Africa. In particular, we proposemultiple features to make a fine evaluation of the degree oflength contrast under different factors such as: read vs semi-spontaneous speech ; standard vs dialectal Wolof. Our mea-sures made fully automatically on more than 20k vowel tokensshow that our proposed features can highlight different degreesof contrast for each vowel considered. We notably show thatcontrast is weaker in semi-spontaneous speech and in a nonstandard semi-spontaneous dialect. Index Terms : computational phonetics, vowel length contrast,automatic speech recognition, wolof language, under-resourcedlanguages
1. Introduction
Growing digital archives and improving algorithms for auto-matic analysis of text and speech create new research oppor-tunities for fundamental research in linguistics and phonetics.This vision is shared by [1] where audiobooks (large amount ofrecordings in many languages and dialects, distributed in a nat-ural way across a wide variety of speakers) are used for corpus-based phonetics. In their work, authors claim that - for the pho-netic events observed - “the data used from audiobooks offersmore tokens than have been examined in the entire 50-year his-tory of sociolinguistic study of Spanish” . In a similar trend, wehave recently shown the value of stochastic and neural acousticmodels for analyzing, at a relatively large scale, vowel lengthcontrast in two under-resourced african languages [2]. Such em-pirical approaches allow statistical evaluation of a much largerset of hypothesis about phonetic variation and its conditioningfactors (among them geographical / dialectal variants). This pa-per illustrates this vision and proposes a detailed analysis ofvowel length constrast in Wolof under different factors such as:read vs semi-spontaneous speech ; standard (Dakar) Wolof vs dialectal (Faana-Faana) Wolof. Paper contributions.
The first contribution of this paperis a large scale analysis of vowel length contrast on Wolof readspeech. Multiple features are proposed to judge the degreeof bimodality in the distribution (of durations) for a givenvowel. Our measures made on 14k vowel tokens show differentdegrees of contrast according to the vowel considered. We alsoshow, in a second contribution, that in the case of read speech,the need of manual transcriptions can be relaxed since the use of automatic speech recognition (ASR) can lead to very similarmeasurements and to the same conclusions. Our third contri-bution is an application of our machine-assisted methodologyto study vowel length contrast in more spontaneous speech forWolof and for one of its dialectal variant (Faana-Faana). Forreproductible research, a Wolof ASR VM and the data of thisstudy are also made available online . Languages studied.
Wolof is the vehicular language ofSenegambia (Senegal and Gambia), also spoken in Mauritania.This paper focuses on senegalese Wolof. We will use the term“standard” to refer to Wolof spoken in Dakar by native speak-ers of the language and “urban” for Wolof spoken by non-nativespeakers. In Senegal, there are also dialectal variants but mutualunderstanding exists between people living in the different ar-eas. Linguists observe some phonetic or morpho-phonologicalvariations, focusing on vocalism, on some forms of verbal in-flection [3] and also on some morphological and syntacticalvariations [4], [5].The Faana-Faana dialect studied in this paper is spoken inthe region of Kaolack, also named
Wolof of the Saloum . It isdescribed by Dram´e [6] and is closer to the Wolof of Gambia.This regional variant is not much influenced by other Wolof di-alects. However, young people and men often spend part oftheir lives in Dakar and come back with influences from stan-dard Wolof. Faana-Faana speakers live in a predominant Sereerspeaking area which influences their own language, but they arenot subject to other major linguistic influences.In Wolof, the vocalic system is composed of 8 short vowels/i/, /e/, /E/ , /a/, /@/ , /O/ , /o/, /u/; each having a long counterpart(except /@/ ). There is no tone in Wolof but phonemes can varyin length [7]. This means that word sense may differ dependingon phoneme duration. For instance, the pronunciation of “fit”( bravery ) and “fiit” ( trap ) varies only at the vowel length level,as well as “wall” ( to rescue ) and “waal” ( to take advantageof ), or “set” ( to be clean ) and “seet” ( to look for ). Same shortand long vowels exists in the Faana-Faana variant. As canbe seen in the examples above, reduplication of the vowel, inthe spelling of Wolof, encodes the duration. One goal of thispaper is to verify if this expected (phonological) contrast is alsoobserved at the phonetic level. Paper outline.
This paper is organized as following. Sec-tion 2 reviews previous works on phonemic contrast analysis.In Section 3, we propose several features to measure degree of(length) contrast for a given unit. In Section 4, we present ourmulti factor analysis of vowel length contrast in Wolof read andsemi-spontaneous speech (Dakar and Faana-Faana). Finally,Section 5 concludes this work and gives some perspectives. see https://github.com/besacier/ALFFA_PUBLIC/blob/master/ASR/WOLOF/WOLOF-VM/ and https://github.com/besacier/ALFFA_PUBLIC/blob/master/ASR/WOLOF/INTERSPEECH_2017 a r X i v : . [ c s . C L ] J un . Related Works Vowel duration is a phonetic measure widely used in speechacoustic research. Many factors affect vowel duration such asits location within the vowel space ([8], [9]), position and lengthof the word [10], surrounding context of the vowel ([11], [12]),speech rate ([13], [14]) and position of the vowel within theword [15]. As raised by [16], main past studies of vowel dura-tion were done through manual annotations. It is consequentlya very time-consuming task and only few words were generallyanalyzed. We believe that use of automatic tools can lead tomore objective and reproductible measures, at a larger scale.As far as vowel length contrast is concerned, [17] studiedits production and perception in Korean. They found that allKorean speakers of the study produced (length) contrasted vow-els but they also concluded that short/long contrast is weaker inspontaneous speech. Vowel length contrast was also investi-gated to better understand language acquisition. [18] analyzed11 hours of Japanese infant-directed speech, using statisticalmethods, to explore how infants learn to discriminate vowellength contrast existing in Japanese. They discovered that dura-tion distribution for a given vowel is not clearly bi-modal sincelong vowels may be much less frequent than short vowels.In Wolof, very few phonetic studies were published, espe-cially on vowel length contrast. One exception is the work of[19] who studied a dialectal variant of Gambian Wolof, closeto Faana-Faana analyzed in this paper. The author compared3 minimal pairs, each containing /i/, /a/ and /u/ vowels (readspeech) and noticed that length contrast was more important forvowel /a/ than for /i/ and /u/. Moreover, less (length) contrastwas observed in rapid speech rate compared to normal speechrate. Finally, in 2006, [7] pointed out that a large analysis ofWolof phonetics was lacking and to the best of our knowledgethis is still the case at present.
3. Measuring Vowel Length Contrast
It is not trivial to objectively analyze the degree of bimodalityin the distribution of durations for a given vowel. One reasonis that - for some vowels - there may be much more short oc-curences than long ones [20]. Eye-looking at distributions is apossibility but more objective features are needed if we want afine evaluation of the degree of contrast across different speechstyles and dialects (see [18] for Japanese). This section pro-poses different criteria (features) to estimate the degree of bi-modality for the (duration) distribution of a given vowel. Thesefeatures are not extracted from true distributions of short andlong vowels, but from their normalized gamma approximations - see Figure 1 for the notations used: (1) ratio r , (2) ratio r ,(3) area A between both (short/long) gamma distributions and(4) delta ∆ between modes of both gamma distributions.We define d S ( x ) and d L ( x ) as representing respectively thedistribution of the short and long units of a vowel (for instance d /i/ ( x ) and d /ii/ ( x ) ). In accordance with this definition, r is defined by equation (1) and is the ratio between d S ( a ) and d L ( a ) , when a is the global maximum value of d S ( x ) . A highvalue of r means a large amount of short tokens compared tolong tokens at the maximum peak of d S ( x ) . In the same way, r defined in equation (2) is the ratio between d L ( b ) and d S ( b ) ,when b is the global maximum value of d L ( x ) . A high valueof r means a large amount of long tokens compared to shorttokens at the maximum peak of d L ( x ) . For both ratios, thebigger the value, the stronger the duration contrast is. We preferred
Gamma distributions to
Gaussian for their skewness. r = d S ( a ) d L ( a ) (1)where a = arg max x ( d S ( x )) . r = d L ( b ) d S ( b ) (2)where b = arg max x ( d L ( x )) . A corresponds to the computed area between both curveswhen d S ( x ) < d L ( x ) , as shown in equation (3). The larger thearea, the stronger the duration contrast should be. We considerthat a significant contrast should give an area A > . . A = (cid:90) ∞ I d L ( x ) − d S ( x ) d x (3)We also compute ∆ which is the difference between bothmodes of d S ( x ) and d L ( x ) , as represented in equation (4). Thegreater the value of ∆ , the more significant the contrast is. Fig-ure 1 displays duration histograms, associated gamma curvesand notations, for phoneme /a/. ∆ = arg max x ( d L ( x )) − arg max x ( d S ( x )) (4)Finally, it is important to note that we did not use Harti-gan’s Dip test of unimodality [21] since our preliminary mea-surements have shown that this test always concludes to the bi-modality of our distribution - even for extremely weak contrasts.
4. Machine Assisted Analysis of VowelLength Contrasts in Wolof
In addition to our existing in-house (Dakar standard) Wolof readspeech corpus [22], we recently collected data during a fieldtrip in Senegal.We collected semi-spontaneous speech of Wolof(Dakar standard) and dialectal variants. In total, we gatheredaround 1.5 hours of elicitated speech from 22 speakers (6 Faana-Faana speakers, 2 Lebu speakers, 3 speakers of urban Wolofand 11 speakers of standard Wolof). Each speaker had to watcha series of 76 short videos designed to express trajectory [23].This data can be considered as semi-spontaneous speech.Our best Wolof ASR system was used to decode newrecorded speech. This is a standard context dependent DNN-HMM hybrid system trained with Kaldi speech recognitiontoolkit [24]. More details on this system can be found in [2]and it is made available through a VM . We used 5 transcrip-tions of Faana-Faana (over 6) and 3 transcriptions of standardWolof (over 11), because only a subset of ASR hypotheses werecorrected by Wolof linguists. Table 1 summarizes each data seton which we will measure vowel length contrast in this paper.Table 1: Wolof speech data overview.
Data Set Male Female
Wolof (read) 8 6 1,120 10,461 1h12 minsWolof (semi-spontaneous) 2 1 254 2,825 14 minsFaana-Faana (semi-spontaneous) 5 0 454 3,365 19 mins see https://github.com/besacier/ALFFA_PUBLIC/blob/master/ASR/WOLOF/WOLOF-VM/ .2. Analysis on Wolof Read Speech In a first phase, we extract vowel durations by force-aligninghuman transcriptions of development ( dev ) set described in[2] (1,120 utterances, 1h12mn of speech) and made up ofWolof read speech (see Table 1). Forced-alignment is donewith our CD-DNN-HMM-based acoustic model (length con-trasted acoustic models with different units for short and longvowels). The 7 contrasted vowels are tagged as /short/ or/long/ depending on the duplication of the grapheme withinthe word. Data is partitioned in different sets denoted by D vl where v ∈ V = { i, e, E , a, O , o, u } is the studied vowel and l ∈ L = { S, L } is the expected length of the vowel (short orlong). We computed vowel durations and built their histogramfor each vowel after deleting outliers (we keep observations x sothat µ − σ < x < µ + 3 σ ). We also approximate our real dis-tribution by the probability density function of a Gamma distri-bution. Eye-looking at normalized distributions for each vowelconfirms that bimodality exists for all of them. However, thedegree of contrast differs for each vowel. For instance, strongduration contrast is observed for vowel /a/ (Figure 1) whereasweak contrast is observed for vowel /O/ (Figure 2).Figure 1:
Histogram and Gamma Distribution for vowel /a/ inWolof Read Speech - Strong Contrast short /O/long /O/
Figure 2:
Histogram and Gamma Distribution for vowel /O/ inWolof Read Speech - Weak Contrast
Table 2 shows measurements of length contrast. Vowels aresorted according to their height. In addition to the contrast fea-tures described in Section 3, we also display in third column themean duration µ (in ms) for each short and long vowel. Vowel/a/ is the one that appears most frequently (both short and long)while vowel /o/ is the one that appears most rarely. This is eas-ily explained because words containing the vowel /a/ are verycommon while those containing vowel /o/ are rare in Wolof.We observe that 2 articulatory features affect vowel duration:height and backness. Indeed, mean duration of short vowels in- Table 2: Contrast Features Extracted on Wolof Read Speech.
Phoneme µ r r A ∆ shortlong (in ms) (in ms) /i/ /i:/
133 131 /e/
227 79 /e:/
178 120 /E/ /E:/
557 131 /a/ /a:/
880 125 /O/
881 73 /O:/
710 102 /o/
60 68 /o:/
69 108 /u/ /u:/
111 110 creases with the aperture of the jaw, as described in [19], exceptfor /a/. The phonological status of /a/ is still in debate and [7]raises the fact that linguists are not all unanimous on the issue.The same rule is not observed on long vowels. Mean durationalso shows that back vowels ( /O/ , /o/ and /u/) are shorter thanfront vowels (/i/, /e/, /E/ ), for both short and long phonemes. ∆ varies from 24 ms to 50 ms and A from 0.27 to 0.56. Vowel/a/ is the one with the strongest length contrast, with large r and r ratios, as well as large area A and large ∆ . Though /O/ is the vowel with the least distinguishable length contrast, withlow r and r ratios, small A and moderate ∆ , features unveilthat all vowels are length-contrasted. The table also shows thatcontrast features are correlated but they are complementary todescribe the shape of the vowel length distributions. To con-clude on this sub-section, this analysis (made fully automati-cally on 14k vowel tokens) show that our proposed features canhighlight different degrees of contrast for each vowel consid-ered and confirm - at a larger scale - previous analyses made. In this sub-section, we try to see if manual transcriptionscan be replaced by ASR hypotheses while keeping sametrends/conclusions. In that case, we relax the constraint of hav-ing manual transcription of the data set. We computed voweldurations from forced alignment obtained with ASR transcripts(from our baseline Wolof ASR system, trained on held-out data- around 20% WER on read speech) and built gamma distri-butions as in previous section. For each vowel, we comparedboth distributions (manual transcription vs ASR transcription)using Kolmogorov-Smirnov statistical test [25] (the null hy-pothesis H was that both distributions obtained after manualand ASR transcriptions are similar). For each vowel v , no sig-nificant difference was found. To illustrate this result, Figure3 shows duration histograms and associated gamma curves forphoneme /u/ when human ( ref ) or ASR ( hyp ) transcriptionsare used for forced-alignment. Both curves are very similar andthis confirms that, for read speech, the need of manual transcrip-tions can be relaxed since the use of ASR leads to very similarmeasurements and to the same conclusions. For the next sub-sections (semi-spontaneous speech), ASR will be also used toproduce transcripts but they will be further corrected by humansdue to the more spontaneous nature of the data . Preliminary measurements have shown that the ASR transcriptionson spontaneous speech are too noisy to be used directly. We got around31% WER for Wolof and 66% WER for Faana-Faana.
50 100 150 200 250 300 duration (in ms) short /u/ of reflong /u/ of refshort /u/ of hyplong /u/ of hyp
Figure 3:
Histogram and Gamma Distribution for /u/ in WolofRead Speech - Using Human (ref) or ASR (hyp) Transcripts
Table 3:
Contrast Features Extracted on Wolof Semi-Spontaneous Speech.
Phoneme µ (in ms) r r A ∆ (in ms) shortlong /i/ /i:/
252 83 /e/
161 71 /e:/
213 83 /E/
518 69 /E:/
225 90 /a/ /a:/
324 100 /O/
360 67 /O:/
190 74 /o/
62 51 /o:/
123 89 /u/
755 51 † /u:/
16 96 † The ratio can not be computed because there were no data for the long unitof the phone ( d L ( a ) = 0 ) at point corresponding to the mode of the shortphone distribution. We computed same features shown in Table 2 on our Wolofsemi-spontaneous corpus. Results are presented in Table 3.Looking at the mean duration of the vowels µ , our firstremark is that it is lower in semi-spontaneous speech com-pared to read speech (for both short and long units). Theseconclusions were expected but they confirm that our machine-assisted methodology allows usable measurements at a largerscale. Comparing µ in read and semi-spontaneous context, weobserve that long vowels are the most affected by the speak-ing style, especially front vowels ( /i:/ , /e:/ and /E:/ ), whileshort units are the least impacted among the vowel set. Resultsfor /u/ have to be taken with caution, since we only have 16long occurences, as well as for /o/ ~ /o:/ for which we have lessoccurences compared to other vowels. All computed featuresshow that length contrast on /O/ ~ /O:/ pair is significantly re-duced in semi-spontaneous speech in comparison to what wasobserved in read speech. In addition, the vowel height has nolonger influence on the duration. Theses findings are consistentwith [26] who described that spontaneous speech have an effecton the vowel pronunciation which tends to be more centralizedwhen pronounced shorter. Table 4: Contrast Features Extracted on Faana-Faana Semi-Spontaneous Speech.
Phoneme µ (in ms) r r A ∆ (in ms) shortlong /i/
882 69 /i:/
167 75 /e/
77 74 /e:/
116 83 /E/
197 69 /E:/
176 87 /a/
909 63 /a:/
188 94 /O/
197 63 /O:/
112 68 /o/
24 53 /o:/
50 77 † /u/ is not represented because we do not have enough data for a comparison. We computed same features shown in Table 2 and Table 3 onour Faana-Faana semi-spontaneous corpus (see Table 4).As we can see in Table 4, long vowels /e:/ and /o:/ stillappear more frequently than their short counterpart, as in semi-spontaneous (standard) Wolof. We observe that the durationincreases with vowel height, for front long vowels ( /i:/ , /e:/ , /E:/ ) but not for their short counterparts. By looking at thevalue of the features, we note that distinction between shortand long pronunciation of vowels is tenuous. The length con-trast on vowel /O/ is also weakened, as in semi-spontaneous(standard) Wolof. These results do not allow to demonstratethat there exists in Faana-Faana a strong opposition of vow-els length as observed in (standard) Wolof. In the mean time,we can not affirm that vowel length contrast does not exist inFaana-Faana. In the descriptions of this dialect, as in the Gam-bian Wolof, the short/long opposition is described, so we canhypothesize that dialectal differences in Wolof are not basedon this lack of contrast. In addition, two-sample Kolmogorov-Smirnov tests revealed that /e/, /E/ , /a/, /O/ vowel distributionsin semi-spontaneous Wolof data set were not found significantlydifferent from those in semi-spontaneous Faana-Faana data setbut /i/, /o/ and /u/ vowel distributions were. Finally, since thisvariant has been little studied, we hope that our analysis repre-sent one first stone in the study of phonemic contrast in Wolofdialects.
5. Conclusion
We presented in this study a large scale analysis (compared toprevious phonetic studies) of vowel length contrasts in Wolof.We worked on different speaking styles but also on one dialec-tal variant (Faana-Faana). We proposed correlated but comple-mentary features to describe the shape of the vowel length dis-tributions and to highlight different degrees of length contrastgiven a vowel. Another important result is that relaxing theconstraints on the transcriptions (by using ASR transcriptionsinstead of manual transcriptions) is possible for read speechsince it leads to very similar distributions of durations. Futurework will be dedicated to leveraging computational models andmachine learning for large scale speech analysis and laboratoryphonetics. Further work will analyze the relation between thesedistinctive features of the length contrast distribution and thefunctional load concept developed by [27]. . Acknowledgements
This work was realized in the framework of the French ANRproject ALFFA (ANR-13-BS02-0009).
7. References [1] N. Ryan and M. Liberman, “Large-scale analysis of Spanish /s/-lenition using audiobooks,” in
Proceedings of the 22d Interna-tional Congress on Acoustics , Buenos Aires, Argentina, 2016.[2] E. Gauthier, L. Besacier, and S. Voisin, “Speed perturbation andvowel duration modelling for ASR in Hausa and Wolof lan-guages,” in
Proceedings of Interspeech , San Francisco, California,USA, September 2016 2016.[3] S. Robert, “Le wolof,” in
Dictionnaire des langues , ser. DicosPoche, J. B. . A. P. Emilio Bonvini, Ed. Quadrige/P.U.F., 2011,pp. 23–30. [Online]. Available: https://hal.archives-ouvertes.fr/hal-00600630[4] S. Voisin and M. Dram´e, “Inaccompli et complexe verbal dansdiff´erentes variantes du wolof,”
Africana Linguistica , still in prep.[5] S. Voisin, “Le wolof et ses variantes,”
JWAL , still in prep.[6] M. Dram´e, “Phonologie et morphosyntaxe compar´ees de trois di-alects wolof,” Ph.D. dissertation, UCAD, Dakar., 2012.[7] M. T. Ciss´e, “Probl`emes de phon´etique et de phonologie enwolof,”
Revue ´electronique internationale de sciences du langageSudLangues , vol. 6, pp. 1–41, 2006.[8] B. Lindblom, “Vowel duration and a model of lip mandible co-ordination,”
Speech Transmission Laboratory Quarterly ProgressStatus Report , vol. 4, pp. 1–29, 1967.[9] I. Lehiste,
Suprasegmentals . MIT Press, Cambridge, MA, 1970.[10] B. Lindblom, B. Lyberg, and K. Holmgren,
Durational patternsof Swedish phonology: do they reflect short-term motor memoryprocesses?
Indiana University Linguistics Club, 1981, vol. 3.[11] A. S. House, “On vowel duration in English,”
The Journal of theAcoustical Society of America , vol. 33, no. 9, pp. 1174–1178,1961.[12] I. Maddieson, “Phonetic cues to syllabification,”
UCLA Workingpapers in phonetics , vol. 59, pp. 85–101, 1984.[13] T. Gay, “Mechanisms in the control of speech rate,”
Phonetica ,vol. 38, no. 1-3, pp. 148–158, 1981.[14] H. S. Magen and S. E. Blumstein, “Effects of speaking rate on thevowel length distinction in Korean.”
Journal of Phonetics , no. 21,pp. 387–410, 1993.[15] S. Myers, “Vowel duration and neutralization of vowel length con-trasts in Kinyarwanda,”
Journal of Phonetics , vol. 33, no. 4, pp.427–446, 2005.[16] Y. Adi, J. Keshet, E. Cibelli, E. Gustafson, C. Clopper, andM. Goldrick, “Automatic measurement of vowel duration viastructured prediction,”
The Journal of the Acoustical Society ofAmerica , vol. 140, no. 6, pp. 4517–4527, 2016.[17] G. Lee and D.-J. Shin, “An acoustic and perceptual investigationof the vowel length contrast in Korean,”
Journal of the Koreansociety of speech sciences , vol. 8, no. 1, pp. 37–44, 2016.[18] R. A. Bion, K. Miyazawa, H. Kikuchi, and R. Mazuka, “Learningphonemic vowel length from naturalistic recordings of Japaneseinfant-directed speech,”
PloS one , vol. 8, no. 2, p. e51594, 2013.[19] R. Sock, “L’organisation temporelle de l’opposition de quantit´evocalique en wolof de gambie. sa r´esistivit´e aux conditions dedur´ee segmentales et suprasegmenales.” Ph.D. dissertation, 1983.[20] S. Sauvageot,
Description synchronique d’un dialecte wolof: leparler du Dyolof . Institut Franc¸ais d’Afrique Noire, Dakar.,1965, no. 73.[21] J. A. Hartigan and P. Hartigan, “The dip test of unimodality,”
TheAnnals of Statistics , pp. 70–84, 1985. [22] E. Gauthier, L. Besacier, S. Voisin, M. Melese, and U. P. Elin-gui, “Collecting Resources in Sub-Saharan African Languages forAutomatic Speech Recognition: a Case Study of Wolof,”
LREC ,2016.[23] C. Grinevald, “On constructing a working typology of the expres-sion of path,”
Faits de langues , no. 3, pp. 43–70, 2011.[24] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek,N. Goel, M. Hannemann, P. Motl´ıˇcek, Y. Qian, P. Schwarz et al. ,“The kaldi speech recognition toolkit,” 2011.[25] F. J. Massey Jr, “The Kolmogorov-Smirnov test for goodness offit,”
Journal of the American statistical Association , vol. 46, no.253, pp. 68–78, 1951.[26] C. Gendrot and M. Adda-Decker, “Impact of duration on F1/F2formant values of oral vowels: an automatic analysis of largebroadcast news corpora in French and German,”
Variations , vol. 2,no. 22.5, pp. 2–4, 2005.[27] E. Ferragne, N. Bedoin, V. Boulenger, and F. Pellegrino,“The perception of a derived contrast in Scottish English,”in