Multilingual Speech Data Collection for the Assessment of Pronunciation and Prosody in a Language Learning System

Transcription

1 Multilingual Speech Data Collection for the Assessment of Pronunciation and Prosody in a Language Learning System O. Jokisch 1, A. Wagner 2, R. Sabo 3, R. Jäckel 1, N. Cylwik 2, M. Rusko 3, A. Ronzhin 4, R. Hoffmann 1 (1) TU Dresden, Laboratory of Acoustics and Speech Communication, Dresden (2) Adam Mickiewicz University, Institute of Linguistics, Poznan (3) Slovak Academy of Sciences, Institute of Informatics, Bratislava (4) Russian Academy of Sciences, Institute for Informatics and Automation, Petersburg oliver.jokisch@tu-dresden.de wagner@amu.edu.pl milan.rusko@savba.sk ronzhin@iias.spb.su Abstract Speech data design and collection are essential tasks for the development of speech technology. The article describes methods, procedures and some results of the multilingual data collection in the EURONOUNCE project. The main objective of the project is the development of a language tutoring system for the pronunciation training of the Slavonic target languages Polish, Slovak, Czech and Russian. To survey language interferences, the example scenario assumes language learners with German mother tongue. Vice versa, the project examines Slavonic mother tongue speakers learning German. The database includes different parts like a reference database of mother tongue speakers by involving source and target languages, speech utterances of language learners at different proficiency levels, training data for the integrated automatic speech recognizer and baseline exercises for the tutoring system. The current speech database contains more than 130 speakers in 5 languages. It represents ca. 200 hours of gross speech data recorded in a studio or in a quiet room. The preliminary results indicate that typological contrasts are dominating as accent source for German learners of Slavonic languages as well as a challenge for students (with a Slavonic mother tongue) learning German. Nevertheless, the peculiarities of non-native pronunciation are specific for given language pairs and depend on both language structures. 1. Introduction The methods of computer-assisted language learning (CALL) and so-called intelligent language tutoring systems (ILTS) play an increasing role in the second language education. The ILTS system, used in the EURONOUNCE project (Jokisch, Cylwik et al [1][2]), provides speech signal analysis functions on the users speech input which allows the user to compare his pronunciation with the tutors one and to receive selective information about potential improvements of articulation. The speech signal processing involves speech recognition and feedback technologies using signal analysis. This article describes the multilingual speech data development as a core task of the project. The chosen sets of source (native) languages and the target (taught) languages includes widely-used languages like German and Russian and national languages which are less taught as foreign languages, like Polish, Czech and Slovak. The project supports teaching and private studies of languages in neighbor countries, e. g. using e-learning infrastructure and computer-based language courses. Pronunciation difficulties represent one of the main reasons that the number of German students learning West Slavonic languages is very low. Computer-assisted language learning methods enable creative approaches in the field of language teaching. Additionally, the implementation of information technology in a multilingual language tutoring system augments the interests of young people who are familiar with computers and communication devices. The EURONOUNCE project, its scope and partners are briefly described in the section 2. The section 3 deals with the multilingual database development, and the section 4 gives an overview about a preliminary analysis and first results of the database development. 2. The EURONOUNCE project The project started November 2007 and is going to be finished in October EURONOUNCE is also intended to build an interdisciplinary and multinational network of teachers, linguists, phoneticians, speech technologists and experts for dissemination. The research and development has focused on Slavonic languages so far and the cross-lingual effects on German or vice versa EURONOUNCE baseline system: AzAR The baseline platform AzAR (German acronym for automat for accent reduction ) was developed in preceding projects (Jokisch, Jäckel et al [3][4][5]). The core function is based on different phonetic-phonologic and prosodic distance measures, involving typical cross-lingual influences from a native source language on the target language. It leads to the marking of mispronounced phones within a spoken utterance using a coloured scale from red ( very bad ) to green ( very good ). It is using confidence measures from a HMM based speech recognizer. The didactic content follows real lessons for the pronunciation training with contrastive exercises, insertion tests, etc. The software was originally developed for Russian migrants in Dresden learning German Partners and Tasks The development consortium consists of TU Dresden (Germany), Adam Mickiewicz University in Poznan (Poland), Slovak Academy of Sciences in Bratislava, voice INTER connect GmbH in Dresden and Russian Academy of Sciences

2 in St. Petersburg. The technically oriented EURONOUNCE tasks of these partners consider following basic issues: Specification of speech data, algorithms and tools; Recording and annotation of multilingual speech databases for L2 tutoring and for ASR; Data analysis and speech signal processing; Integration into tutoring and courseware systems; Technical support for evaluation and dissemination. Acknowledged partners from language education, e.g. the Goethe-Institut, evaluate the systems concept in real lessons. 3. Multilingual Speech Databases 3.1. Requirements of pronunciation assessment It is widely recognized that pronunciation errors are specific for certain pairs of native/target languages (Flege & Wang [6], Kenworthy [7], Rochet [8]). Pronunciation difficulties arise due to differences or similarities of phonetic structures [8]. Learners produce pattern structures that differ from normative patterns of the target language in a symptomatic way depending on their native language L1. Starting with the second language (L2) training, the learners sound system is not yet organized as L2 normative speech. Phonemes or allophones are often replaced by similar phonemes, regarding the viewpoint of the learner. Failure patterns remind of transliteration rules if different alphabets or spellings are used in L1 and L2 (e. g. replacement of the glottal fricative /h/ by /g/, /x/ or /x / in Russian accented German or replacement of /i/ following a non-palatalized consonant by /y:/ or /Y/ in German accented Czech). Automatic pronunciation assessment can be based (1) on a projection of transfers or interferences onto a system of linguistic categories, e. g. onto the IPA categories for the determination of phonologic feature distances [9] or (2) on acoustic parameters [10]. Both approaches allow for the computation of segmental distances and Levenshtein distances between a standard representation and a non-native pronunciation variant and, consequently, for the computation of transformation costs. The learning progress in a scope of pronunciation training can be defined as minimization of transformation costs between a target and learners pattern. However, not all of the transfers can be mapped on the segmental level. For instance, if a learner produces a pattern structure of an L2 utterance /ABCD/ using the string /ABCF/ where /D/ is not a phonemic unit of the learners L1, and /F/ is an element of the phoneme inventories of both, L1 and L2, /ABCF/ can be either well-formed or not. If there is a constraint in L2 that forbids /F/ to occur in a final position, /ABCF/ will not be well-formed. It is necessary to distinguish between substitutions of segments that violate a phonotactic constraint and those that do not. Usually, phonologic categories, distinctive features and correlations between classes of segmental units are not identical in L1 and L2. In many cases, transfers or interferences can be attributed to incorrect classifications. Attention is given to unequal categorizations that represent typological differences. For instance, German belongs to a class of languages with duration and tenseness contrast within the vowel system and a syllable section contrast (German Silbenschnittkontrast ) while Russian does not have these quantitative correlations. The construction of the text base for the speech recordings is based on a contrastive analysis of inventories, phonologic correlations and phonotactic constraints as well as on the analysis of different relations between prosodic patterns and lexical, morphologic and syntactic structures of L1 and L Methods of database design The collection and the type of text material for the recordings was dictated by a specific application of the resulting speech material. Generally, six tests including different types of texts were proposed. The accent and dialectological tests serve the purpose of investigating L1 interferences and consist of sentences created by experienced teachers of phonetics. The dialectological test also includes words/sentences for which alternative pronunciations exist. The Phondat test contains text material designed for the purpose of ASR training (recognition of non-standard pronunciation), analysis and assessment of pronunciation errors. The sentences were selected mostly from existing text corpora designed for speech applications (TTS and ASR) and modified if necessary. The resulting sentence sets are appropriate for a specific proficiency level. A part of the text corpus contains phrases for the fluent reading test which aims at the analysis of the realization of segmental, but also suprasegmental and discourse features of non-native speech. The spontaneous speech test is used to collect information on speakers proficiency level, but the main objective of the test is to investigate different aspects of spontaneous speech produced by non-natives. This test is addressed to advanced students only. The prosodic test contains text material (phrase and short dialogues) created by linguists for investigating the realization of prosodic/intonational features by advanced L2 learners and L1 interferences in the domain of prosody Database design and structure The multilingual speech database is structured according to the language pair, proficiency level A1 - C2 [11], test type and language of the test (L1, L2). Table 1: Multilingual speech database (overview) Level Preliminary (A1 - B1) Advanced (B2 - C2) L1 - L2 Language pair Dialectological test Accent test Phondat (part 1) Text for fluent reading (1) Dialectological test Accent test Phondat (part 1, 2 & 3) Fluent reading (1, 2 & 3) Spontaneous speech test Prosodic test L2 - L1 Accent test Prosodic test Dialectologica l test Phondat (part 1, 2 & 3)

3 Table 1 illustrates the general structure of the multilingual speech database. A part of the text material is read by both native and nonnative speakers, because our objective is to collect native speech data which can serve as a reference for the assessment of non-natives pronunciation. Cross-lingual examples between Polish, Slovak, Russian and German are described. The accent test consists of sentences posing pronunciation problems of a different origin (Bonaventura et al [10]): Pronunciation of non-native sounds, e. g. the realization of Polish palatalized fricatives and affricates as in nieg (/s n ek/) or ziemia (/z emja/) by students with L1 German. Realization of Polish voiced/voiceless fricatives or affricates and plosives as lenis and fortis consonants respectively, whereas no such contrast exists in Polish and the modifications of L2 segments pronunciation are perceptible. Students with L1 Slovak often realized German uvular r as alveolar r. They have problems with labial affricate [pf] and often substitute rounded front vowels to back vowels. Possibility of carrying over the pronunciation regularities from students L1 e. g., the substitution of Schwa for a word-final, unaccented vowel by students with L1 German. In Slovak, for example, diphthong ie should be realized as on-glide while the students with L1 German often pronounced it either as long phoneme /i:/ or off-glide. Overgeneralization of L2 regularities such as choosing a possible pronunciation but in an inappropriate context - e. g. German has voicing neutralization in coda consonants that is similar to Slovak but this process is blocked in word-final positions where the following word starts with a vowel. Hence, speakers with L1 German might pronounce phrases like pod oknom and pot otca as po[t] while it should be po[d] in both cases. The dialectological test contains sentences which cover phonemic inventories of the analyzed languages taking into account all vowels and diphthongs in stressed versus unstressed positions (e. g. PL: bi robi ; SK: kô pakô ) and in contrastive minimal pairs (e. g. PL: ma me; SK: sud sad) as well as all consonants in minimal pairs such as voiced vs. voiceless, trill vs. lateral, geminates vs. single (e. g. PL: pasa passa; SK: poda podda ), etc. The test also includes phonemes which are characteristic for the language in question e. g. Polish nasalized vowels, and Slovak diphthongs ia, ie, iu, ô in different phonetic contexts and positions, or consonantal clusters in different positions in a word (initial, middle and final). The test covers obligatory assimilation processes (devoicing and voicing) within the word (e. g. SK: podpredseda) and at the word boundaries (e.g. SK: od pondelka) as well as phonetic and phonological processes that are facultative i.e. depending on a dialect. For the words which have alternative pronunciations, the transcription of all the variants was provided. The resulting speech material is used to analyze mispronunciations by nonnatives. Therefore, it was carefully designed using basic vocabulary and grammatical structures adapted to elementary-level students so that the number of new items is limited. The Phondat sentences include the most frequent words, syllables and triphones in different contexts (accented/unaccented, phrase/word-initial/medial/final, etc.). In order to investigate the spontaneous speech realization by intermediate and advanced students, two tests were designed. They include tasks consisting in explaining the meaning of a phrase by producing an exemplary sentence containing the phrase (test 1) or describing its meaning in students own words (in this task also proverbs are taken into account, test 2). In both cases, only such phrases are used which are present and popular in students L1. In the second part of the tests, students are asked to answer a couple of questions regarding their everyday life (test1) or to finish given sentences (test2). The resulting speech material is used to assess the proficiency level of the student. For each language pair, three texts for the fluent reading test were selected according to the following criteria: Contemporary language (i.e. no outdated vocabulary or rare grammatical structures), Prosody - different types of sentence modes, discourse structures, expressivity, Universal stories are preferred (e.g., stories by H. Ch. Andersen or Grimm Brothers), High word and triphone coverage, Phonetically rich and balanced sentences. The prosodic test consists of exercises aiming at analyzing/practicing the realization of prosodic features which are (a) commonly taught in a foreign language class and thus, included in typical textbooks, (b) easy to teach (e.g., focus) and (c) important for comprehension. The resulting test contains sentences and mini dialogues which should be realized with different intonational patters characteristic for: Questions: yes/no, wh-, contrastive, suggestive, rhetoric, reported/quoted Commands, requests, warnings, statements of a different structure (compound sentences, listings, etc.) Apart from that, a special set of different types of sentences, for which specific rhythmic patterns exist, were designed such as e.g., proverbs. Another set of sentences was designed for the analysis/training of the realization of different stress patterns and irregular word-stress. Minimal pairs i.e., words which have similar spelling in L1 and L2 but differ with respect to stress position were also taken into account (e.g. Slovak ideál vs. German Ideal or Slovak word psycholingvistika vs. German Psycholinguistik). In order to make comparison of the realization of prosodic features by native and non-native speakers possible, the text material included in the prosodic test was read by both groups of speakers. The resulting multilingual speech database provides an extensive basis for the analysis of different aspects of nonnative speech.

4 3.4. Recording specifications Reference speaker: recording database The recordings of a reference speaker will serve as template utterances for the exercises. The pronunciation and prosody should follow the rules of the standard form of the language being taught. This database consists of the whole set of the reference utterances for a given L1/L2 pair, uttered by two native speakers (male and female) with pleasant voices. The reference speakers are recorded under sound-studio conditions of high standard with practically no perceivable reverberation or background noise. High speech signal quality is important for these recordings, therefore a large diaphragm condenser microphone with professional-quality pre-amplifier and AD converter is the main recording input. (e.g., the Slovak partner uses RØDE K2 microphone, MOTU Traveler digital mixing device and a PC-based hard disk recording system.) A second channel is recorded via head-mounted close-talk microphone suitable for speech recognition purposes. Minimum requirements for the reference speech recording are: 44.1 khz sampling frequency, 16 Bit resolution and PCM with little endian format Development data The development data will be used to study foreign accent effects and learning process. The speakers are recorded in a simple studio or quiet-room environment with acceptable acoustic conditions and nearly no reverberation or background noise. Minimum requirements for the recording chain are: Consumer quality audio interface (e.g. E-MU Tracker Pre USB),reported/quoted, Headset (consumer quality) or head-worn microphone, Table microphone: large diaphragm (3/4 inch or more) condenser microphone. The Slovak recording equipment consists of table microphone RØDE NT3, head-mounted microphone Sennheiser ME3 with Sennheiser MZA 900 P XLR adaptor, audio interface E- MU Tracker Pre USB and the quiet (SSD disk equipped) notebook ASUS eeepc 901 with secondary monitor. The speech is recorded with 16 khz or higher sampling frequency (preferably 44.1 khz), resolution 16 Bit, and stored in PCM with little endian format Training data The training data will serve for the training of a generalpurpose speech recognizer and should contain a large number of speakers recorded with real scenario quality. The speakers can be recorded in a normal office room (normal standard) with acceptable acoustic conditions (Minimal SNR of 35 db is intended). It is proposed to record two channels - one with headset (consumer quality) and second with good quality table microphone. Minimum specification is to record at least one track (headset preferable). The speech is recorded with sampling frequency 16 khz or higher (preferably 44.1 khz), resolution 16 Bit, and stored in PCM with little endian format Implementation work For each language pair, 18 speakers were recorded. The speakers were recruited from students of the target language with a different degree of proficiency (preliminary to advanced). The multilingual speech databases were balanced with respect to gender and proficiency of the speakers. The proficiency ratings were based partly on the information provided by the students, namely they were asked to complete a questionnaire including a self-judgment of their language skills. This information was verified by an expert who listened to a sample recording of students speech produced during a short spontaneous speech production test. At the beginning, test recordings were carried out with representatives of all language pairs and proficiency levels. The objective was to verify whether the text material is appropriate for students at a particular proficiency level, as well as to define explicit instructions for the speakers. Consequently, it was decided that preliminary students (level A1-A2) got access to the full text material before the planned recording session, and were encouraged to study the texts so that they could understand their content, but not to practice the reading of the texts. The speakers were asked to read the texts in a moderate tempo and to avoid disfluency as much as possible. They could and were encouraged to re-record sentences in which major errors occurred e.g. misreading of one or more words. In the text material of the tests designed especially for the analysis of the interferences and assessment of pronunciation errors (mainly accent and dialectological test), the speakers re-recorded also these sentences in which the temporal structure was significantly modified as a result of disfluency or speakers hesitation. Our aim was to avoid acquisition of speech data which could not serve as a reliable basis for the analysis of segmental (and especially vowel) duration contrasts. At the start of the recording session, information concerning the speaker (age, gender, city and country of origin, proficiency level) and the recording session itself (date, time, location, test name) was collected. Depending on the proficiency level and type of the text material, the speakers needed between an hour and two hours to complete the whole recording session. Figure 1: Studio recording session at Adam Mickiewicz University in Poznan

5 The recordings were carried out in a studio or a sound-proof room as described before. Figure 1 shows a typical recording session using the equipment described in sec Preliminary results 4.1. Database realization (overview) Regarding the current status of realization of the speech database for the language pair SK-DE/DE-SK, the SK-DE part was completed. It includes 56.5 hours of speech. For the language pair PL-DE/DE-PL, the PL-DE part of the database is almost finished (missing 4 sessions out of 54). It contains about 54 hours of speech. However, due to unavailability of students of Polish with L1 German, only 30% of the DE-PL part has been recorded so far (i.e., seven speakers, about 11 hours and 30 min. of speech). Regarding RU-DE/DE-RU, more than 50 speakers were recorded and the database verification process has started. The whole speech material was segmented and phonetically transcribed using force alignment. For Polish part, a subset of the speech database (including accent, dialectological and phondat sentences) was manually verified: segment boundaries were adjusted wherever necessary, noise, disfluency and pauses were marked, transcription and annotation of automatically inserted primary and secondary stress markers were checked, and deviations from the canonical pronunciation were marked (insertions, deletions and substitutions of phonemes). Word and triphone statistics were carried out for the phondat sentences and the text material for the fluent reading test (i.e. the part of the database which was collected from existing text resources). The Polish phondat corpus includes altogether 341 sentences including 1,839 different words, 5,024 different inter- and 3,530 intra-word triphones. As regards the texts for the fluent reading test for German students of Polish, they contain sentences including up to 251 different words, 1,167 different inter- and 726 intra-word triphones. The Slovak phondat corpus consists of 300 sentences including 1,370 different words, 2,010 different inter- and 3,268 intra-word triphones. The texts for the fluent reading test of German students of Slovak contain 65 sentences including 382 different words, 759 different inter- and 1,001 intra-word triphones. The preliminary statistics indicate a high coverage of the most frequent words and triphones Analysis and experiments (first results) The preliminary analysis of features for the non-native pronunciation based on the PL-DE corpus confirmed the findings on common pronunciation problems, interferences and errors made by learners of German with L1 Polish presented in the literature (e. g. Morciniec 2008 [13]): Realization of phonologically long German vowels: substitution of the sequence of /ej/ for the tense vowel /e:/ or lax long /E:/ e.g. lesen pronounced as /lejzen/ or später realized as /Spejter/, similarly, substitution of /ju/, /u/ or /i/ for tense, long German /y:/ e.g. Schüler /Sjuler/, Difficulties in pronunciation of phonemes not present in students L1 e.g. /I/ as in Sitz (/zits/): students with L1 Polish will probably substitute Polish vowel /i/ or /y/ for the German short lax /I/; the interference consisting in substitution of /x/ for /C/ not present in Polish phoneme inventory as a matter of fact German students of Polish make the opposite substitution and realize /C/ in contexts in which it occurs is German e.g., ich (them) pronounced as /ic/ instead of /ix/. As regards pronunciation of German students of Polish, the following problems have been observed so far: Substitution of German diphthongs /au/, /ai/, /OY/ for sequences /aw/, /aj/, /oj/, Realization of palatalized consonants e.g. nie (no) /n e/ pronounced as /nje/, Pronunciation of Polish voiced plosives, fricatives and affricates with a weaker tension of the articulators which is characteristic for German lenes consonants, but gives impression of a partial devoicing. The opposite tendency is observed for realization of Polish voiceless plosives, fricatives and affricates which are pronounced with a greater tension of the articulators characteristic for German fortes consonants. Realization of /s / and /z / as palatal /S/ and /Z/ respectively, which results from shifting the place of articulation, Glottal stop and glottalization of word initial vowels and glottalization at morpheme boundaries in contexts where it is used in German: This kind of vocal onset is untypical for Polish in which only soft vocal onset occurs e. g., /#o.sta.tn i#aw.to.bus/ pronounced as /#Qo.sta.tn i#qau.to.bus/. A preliminary analysis of features for the non-native pronunciation based on the SK-DE corpus confirmed following common pronunciation issues, interferences and errors by learners of German: Realization of alveolar r instead of uvular r, e.g. pronunciation of the word brennen as /bren@n/ instead of /bren@n/, realization of endings er (e.g. Bruder, aber, lieber as /bru:der/ /a:ber/ /li:ber/ instead of realization with schwa-r /bru:d@6/ /Qa:b@6/ /li:b@6/). Difficulties in pronunciation of phonemes not present in students L1. In Slovak, it is phoneme /Y/ ü which is pronounced as /i/, e.g. word Müller mispronounced as /miler/ or /mil@6/ and phoneme /2:/ ö mispronounced as /E:/, e.g. das Möbel mispronounced as /das me:bel/. We also observed that non-native speech for both L2 languages (SK and DE) displays decreased pitch range compared to the native speech (see also Gut 2003 [14]). The L1 RU L2 DE corpus has been examined regarding both, qualitative and quantitative deviations from standard German pronunciation [4] [5]. The observed interferences cannot be reduced to particular substitutions. Commonly, they represent profound differences between the phonologic systems: different classes, distinctive features and systematic correlations, different phoneme distribution rules and supra-

6 segmental constraints. Among the perceptively relevant features of Russian accent in German one can observe substitutions of phonemes which do not have a corresponding class in Russian: front rounded vowels /y:/ kühn, /Y/ dünn, /2:/ schön, /9/ köstlich, mid-high vowels /e:/ Gebet and /o:/ Gebot, diphthongs /au/ laut, /ai/ weit, /OY/ deutsch, the glottal fricative /h/ gehören, the velar nasal /N/ lange. Perceptively suspicious substitutions of allophones occur which belong to classes absent in Russian: the unvoiced palatal fricative /C/ (allophone of /x/) Küche and the uvular r- allophones /R/ Gerät, /X/ schroff, /6/ Fahrt. Systematic deviations from standard German pronunciation arise from different phonologic correlations: The German correlation between long tense and short lax vowels (Staat Stadt, Beet Bett, etc.) has no equivalent in Russian. Typical features of a Russian accent in German are lengthening of short vowels (usually under phrase accent) and lacking tenseness of long vowels. On the other hand, accent features are induced by L1- specific categories: Learners with L1 RU frequently substitute German consonants with the corresponding Russian palatal phonemes (e. g. the palatal liquid /l / occurs regularly in combinations with high and front labial vowels and diphthongs /ai/, /OY/ - in words like Flug, Luft, Lüge, Löwe, Löffel, leise, Leute). As a secondary phenomenon, vowels in combinations with Russian palatal consonants underlie the accommodation. Within the subsystem of consonants, main difficulties to learners with L1 RU are caused by the acquisition of the fortis-lenis-contrast. Even advanced learners with L1 RU transfer the Russian voicing correlation and regressive voicing assimilation to German (e. g. hat sich pronounced like */had#zic/). 5. Conclusions The preliminary analysis of learners speech provides evidence for both, analogies and differences between the language pairs involved. The main typological contrasts of German (tense-laxcontrast intrinsically tied to vowel duration and fortis-leniscontrast) are dominating in both directions: as source of German accent in the case of a Slavonic L2 and as special challenges for students with a Slavonic L1 and L2 German (reversibility to a certain extent). Nevertheless, the peculiarities of the non-native pronunciation are specific for a given language pair and depend on structural properties of both, source and target languages. The analysis of speech data within the project will be continued with a main focus on supra-segmental constraints and prosodic patterns. 6. Acknowledgements This project has been funded with support from the European Commission within the Lifelong Learning Programme (project LLP DE-KA2-KA2MP). This publication reflects the views only of the authors, and the Commission cannot be held responsible for any use which may be made of the information contained therein. The project homepage is located at: 7. References [1] Jokisch, O., Jäckel, R., Rusko, M., Demenko, G., Cylwik, N., Ronzhin, A., Hirschfeld, D., Koloska, U., Hanisch, L., Hoffmann, R., The Euronounce Project an intelligent language tutoring system with multimodal feedback functions, roadmap and specification. Proc. 19 th ESSV 2008, Frankfurt/M., pp , September [2] Cylwik, N., Demenko, G., Jokisch, O., Jäckel, R., Rusko, M., Hoffmann, R., Ronzhin, A., Hirschfeld, D., Koloska, U., Hanisch, L., The use of CALL in acquiring foreign language pronunciation and prosody general specifications for Euronounce Project. Proc. SASR 2008, Piechowice (Poland), September [3] Jokisch, O., Koloska, U., Hirschfeld, D., Hoffmann, R., Pronunciation learning and foreign accent reduction by an audiovisual feedback system. Proc. 1 st Intern. Conf. on Affective Computing and Intelligent Interaction (ACII), Beijing, pp , October [4] Jokisch, O., Jäckel, R. and Hoffmann, R., Phoneticphonologic feedback functions for the explicit German pronunciation training. Proc. Workshop on Speech Analysis, Synthesis and Recognition (SASR), Krakow, September [5] Jäckel, R., Jokisch, O. and Hoffmann, R., Evaluation of the Speaker Proficiency in a Pronunciation Tutoring System. Proc. 12 th Intern. Conf. Speech and Computer (SPECOM), Moscow, , October [6] Flege, J. E. and Wang, C., Native-language phonotactic constraints affect how well Chinese subjects perceive the word-final /t/-/d/ contrast. Journal of Phonetics, 17, , [7] Kenworthy, J., Teaching English Pronunciation. Longman, New York, [8] Rochet, B. L., Perception and production of second language speech sounds by adults. In W. Strange (ed.), Speech perception and linguistic experience. Timonium, MD, York Press, [9] Vieregge, W. H., Rietveld, A. C. M. and Jansen, C., A distinctive feature based system for the evaluation of segmental description in Dutch. Proc. 10th International Congress of Phonetic Sciences, Dordrecht, pp , [10] Heeringa, W., Measuring dialect pronunciation differences using levenshtein distance. PhD thesis, University of Groningen, [11] Common European Framework of Reference for Languages. Retrieved on 21st August 2008 from [12] Bonaventura, P., Herron, P. and Menzel, W., Phonetic rules for diagnosis of pronunciation errors. Proc. KONVENS 2000, pp , Ilmenau, October [13] Morciniec N. and Pr dota S. A, Handbook of German pronunciation. Polish Scientific Publishers (PWN), 2008, ISBN: [14] Gut, U., Prosody in second language speech production: the role of the native language. Zeitschrift für Fremdsprachen Lehren und Lernen 32, pp , 2003.