Proceedings Fonetik 2009 - Institutionen för lingvistik

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityError residualFigure 1. Cross-validated regression residuals. Thedata is ordered in an ascending human error withthe circles indicating the residual errors.The resulting regression residuals indicatethat perhaps some nonrandom trend is stillpresent. Some relevant information not includedby the regression model is therefore possiblystill present in the residual errors. It maybe that the used prosodic features do not includethis information or that more coefficientsin the model could be justified.Regression estimate1050−52520151050Cross−validated residuals5 10 15 20 25 30 35 40 45 50 55 60Sample #Scatterplot−50 5 10 15 20 25Human estimateFigure 2. A scatterplot of human errors against thecorresponding cross-validated regression estimates.The solid line shows for reference where a perfectlinear correspondence is located and the dashed lineis a least squares fit of the data.The scatter plot shows a linear dependenceof 0.76 between human and regression estimateswith 66% of variance explained.ConclusionThe results suggest that segmental fluency or“correctness” in second language speech can bemodelled using prosodic features only. It seemsthat segmental and supra-segmental second languagespeech skills are interrelated. Parametersdescribing the dynamics of prosody (notably,the steepness and magnitude of f0 movements –see Table 1) are strongly correlated with theevaluated segmental quality of the second languagespeech data. Generally, it may be thecase that segmental and supra-segmental (prosodic,intonational) problems in second languagespeech occur together: a command of onepronunciation aspect may improve the other.Some investigators actually argue that good intonationand rhythm in a second language will,almost automatically, lead to good segmentalfeatures (Pennington, 1989). From a technologicalviewpoint, it can be concluded that amodel capable of estimating segmental errorscan be constructed using prosodic features. Furtherresearch is required to evaluate if a robusttest and index of speech proficiency can beconstructed. Such an objective measure can beseen as a speech technology application of greatinterest.ReferencesAhmadi, S. & Spanias, A.S. (1999) Cepstrumbased pitch detection using a new statisticalV/UV classification algorithm. IEEE Transactionon Speech and Audio Processing 7(3), 333–338.Hubert, M., Rousseeuw, P.J., Vanden Branden,K. (2005) ROBPCA: a new approach to robustprincipal component analysis. Technometrics47, 64–79.Khuri, A.I. (2003) Advanced Calculus withApplications in Statistics, Second Edition.Wiley, Inc., New York, NY.Morris-Wilson, I. (1992) English segmentalphonetics for Finns. Loimaa: Finn Lectura.Pennington, M.C. (1989) Teaching pronunciationfrom the top down. RELC Journal, 20-38.Pudil, P., Novovičová, J. & Kittler J. (1994)Floating search methods in feature selection.Pattern Recognition Letters 15 (11),1119–1125.Seppänen, T., Väyrynen, E. & Toivanen, J.(2003). Prosody-based classification ofemotions in spoken Finnish. Proceedings ofthe 8th European Conference on SpeechCommunication and TechnologyEUROSPEECH-2003 (Geneva, Switzerland),717–720.Titze, I.R. & Haixiang, L. (1993) Comparisonof f0 extraction methods for high-precisionvoice perturbation measurements. Journal ofSpeech and Hearing Research 36, 1120–1133.119