Studying Luxembourgish Phonetics Via Multilingual Forced Alignments
Total Page：16
File Type：pdf, Size：1020Kb
ICPhS XVII Regular Session Hong Kong, 17-21 August 2011 STUDYING LUXEMBOURGISH PHONETICS VIA MULTILINGUAL FORCED ALIGNMENTS M. Adda-Deckera,b, L. Lamelb & N.D. Snoerenb aLPP CNRS / Univ. Paris 3, F-75005 Paris, France; bLIMSI CNRS, F-91403 Orsay, France [email protected]; [email protected]; [email protected]; [email protected] ABSTRACT language, due to the fact that the written production is low, and linguistic knowledge and resources, such as Luxembourgish, a Germanic-Franconian language, is lexica and pronunciation dictionaries, are sparse. embedded in a multilingual context on the divide Written Luxembourgish is not systematically taught to between Romance and Germanic cultures and children in primary school: German is usually the first remains one of Europe’s under-described languages. written language learned, followed by French. This paper investigates the similarity between This paper aims to gain more insight into the Luxembourgish phone segments with German, acoustic properties that define the Luxembourgish French and English via forced speech alignment language in the light of its Germanic and Romance techniques. Making use of monolingual acoustic seed influences. In particular the goal is to determine the models from these three languages, as well as influence of the German, French and the less practiced “multilingual” models trained on pooled speech data English language on the acoustic realization of we investigated whether Luxembourgish was Luxembourgish phonemes. The focus is on acoustic globally better represented by one of the individual modeling whilst making use of phonemically aligned languages or by the multilingual model. Although audio data. The following questions are addressed. First, French words are often interspersed in spoken if multiple monolingual acoustic models are available Luxembourgish, forced alignments show a clear for alignment of Luxembourgish audio data, is there a preference for Germanic acoustic models, with only a clear preference for one of the languages? Second, is a limited usage of the French ones. While globally, the language- specific preference observed for specific German models provide the best match, a phone- phonemes or for phoneme classes? If so, how do they based analysis, shows language-specific preferences: correspond to IPA symbol matches between the French is preferred for rounded front vowels, nasal preferred language and Luxembourgish phonemes? vowels and / / whereas English is more frequently Third, how do language-specific preferences, if any, used for diph- thongs. The proposed method enables fare when compared with multilingual phone models? the acoustic match between phonemes in different These issues have important implications for acoustic languages to be quantified and opens new model for automatic speech recognition and handling perspectives in language processing studies for low pronunciation variants. The next section introduces the e-resourced languages and for L2 learning. phonemic inventory of Luxembourgish and its Keywords: forced alignment, acoustic modeling, correspondance with the three source languages multilingual models, Luxembourgish, Germanic (German, French, and English). Results are then languages, Romance languages presented with both monolingual and multilingual seed models. Finally, section 6 summarizes the results and 1. INTRODUCTION discusses future challenges for speech technology and Luxembourg, a small country of less than 500,000 linguistic studies of Luxembourgish. inhabitants in the center of Western Europe, is com- 2. PHONEMIC INVENTORY OF posed of about 65% of native inhabitants and 35% of LUXEMBOURGISH immigrants. The national language, Luxembourgish ("Lëtzebuergesch"), has only been considered as an The adopted Luxembourgish phonemic inventory official language since 1984 and is spoken by includes a total of 60 phonemic symbols including 3 natives.[4]. The immigrant population generally extra-phonemic symbols (for silence, breath and speaks one of Luxembourg’s other official languages: hesitations). Table 1 presents a selection of the French or German. Recently, English has joined the phonemic inventory together with illustrating set of languages of communication, mainly in examples (see [4] for more information. on the professional environments. phonemic inventory of Luxembourgish). As pointed out by [2] and [3], Luxembourgish Luxembourgish is characterized by a particularly should be considered as a partially under-resourced high number of diphthongs. 196 ICPhS XVII Regular Session Hong Kong, 17-21 August 2011 Table 1: Sample cross-lingual phoneme mappings: Lux. targets mapped to same or similar phonemes in 3 source lang. (Fre, Ger, Eng). Table 2: Phoneme and training data information (in hours) for native and pseudo-Lux. acoustic models from English, French, German and for the multilingual superset (E, F, G). To minimize the phonemic inventory size, we could have chosen to code diphthongs using two consecutive symbols, one for the nucleus and one for the offglide (e.g. the sequence /a/ and /j/ for diphthong ). We prefered, however, the option of Table 1 shows a sample of the adopted cross- coding diphthongs and affricates using specific lingual mappings that were used to initialize seed unique symbols. Given the importance of French models for Luxembourgish. Some symbols are used imports, nasal vowels were included in the inventory, several times for different Luxembourgish phonemes. although they are not required for typical For the diphthongs that are missing in French, Luxembourgish words. Furthermore, native phonemes corresponding to the nucleus vowel were Luxembourgish makes use of a rather complex set of chosen. A fourth model set was then formed by voiced/unvoiced fricatives. The final Luxembourgish concatenating the first three model sets, allowing the inventory includes 60 symbols. decoder to choose among the three models (see Table 2). Finally a set of multilingual acoustic models were 3. ACOUSTIC SEED MODELS trained (see Fig. 1, lower part) using the pooled The need to develop acoustic seed models for un- E,F,G audio data that were labeled using their derresourced languages has already been addressed in respective IPA(i,n) correspondances. previous research [5]. In the current study, three sets 4. LUXEMBOURGISH SPEECH DATA of context-independent acoustic models were built, one for each well-resources seed language (i.e., Forced alignments were carried using 80 minutes of English, French, German). The models were trained manually transcribed speech from the House of on manually transcribed audio data (between 40 and Parliament (Chamber debates (70’) and from news 150 hours) from a variety of sources, using language- (10’) broadcast by RTL, the Luxembourgish radio specific phone sets. The amount of data used to train and TV broadcast company [2]. The detailed manual the native acoustic models and the number of transcripts include all audible speech events, includ- phonemes per language are given in Table 2 (left). ing disfluencies and speech errors. These verbatim Each phone model is a tied-state left-to-right, 3-state transcripts were checked against the resulting word CDHMM with Gaussian mixture observation lists for errors and orthographic inconsistencies. densities (typically containing 64 components). The average Luxembourgish phone segment du- Figure 1 (top) illustrates the development of three ration remains relatively stable with respect to the sets of pseudo-Luxembourgish acoustic models, each different seed alignment (70-80ms). The corpus in- including 60 phones, starting from the English, cludes a total of 56,000 phone segments. French and German seed models and mapping the Luxembourgish phonemes to a close equivalent in 5. RESULTS each of the source model sets (IPA(i,n) in Fig. 1). In our earlier investigations [1], a language change in acoustic models was permitted only on word boundaries. In that case, a clear preference for the” 197 ICPhS XVII Regular Session Hong Kong, 17-21 August 2011 German” acoustic models could be noticed with more phoneme repertoires, there were many shared IPA than 50% of phone segments aligned with the symbols (e.g. plosives). For some of the symbols (e.g. acoustic models arising from the German audio data. diphthongs) however, a correspondance was fixed In the present study, the language identity of the approximatively (i.e., if the diphthong does not exist acoustic models may change at any phone segment in the language of the audio training data, the boundary. As a result, a large number of language corresponding nucleus vowel was assigned to the switches between model sets are observed. The Luxembourgish diphthong). In the following, we German models are less used. However, in all investigate the alignment results as a function of the explored alignment conditions (three sets of models closeness of the IPA match. trained with monolingual English, French and For each phoneme symbol, all observed segments German data, or adding also a fourth acoustic model were aligned with one of the provided acoustic model set trained with the pooled audio data), the German sets (En, Fr, Ge, Pooled). Rates were computed per models are always at rank one. In the alignment language origin of the acoustic model set. When not condition including the pooled model set, German considering the pooled models, the English, French models are aligned with 36% of segments, followed and German rates sum up to 100% for each symbol by the English set (27%) and the pooled model set (results