CrossTowns: Automatically generated phonetic lexicons
of cross-lingual pronunciation variants of European city names.
This is a (lengthy) abstract of my article:
Schaden, S. (2004): "CrossTowns: Automatically generated phonetic lexicons
of cross-lingual pronunciation variants of European city names." Proceedings Fourth International Conference on Language Resources and Evaluation (LREC 2004). Lisbon, Portugal, 1395-1398.
The pronunciation of place names (toponyms) by non-native speakers is a problematic issue for speech technology applications such as travel information or car navigation systems: In this application scenario, a broad spectrum of potential mispronunciations and pronunciation variants must be anticipated, ranging from minor phonetic shifts to strongly accented forms that hardly resemble the canonical forms provided in ASR and TTS phonetic dictionaries. This mismatch between expected pronunciation and actual pronunciation may lead to a significant performance decrease for automatic speech recognition and synthesis in terms of recognition rate for ASR (cf. van Compernolle, 1999), or naturalness and acceptability for TTS (cf. Dahlbäck et al. 2001).
The CrossTowns pronunciation lexicon project is based on previous work introduced by the author at LREC 2002, where a speech database for the analysis of cross-lingual pronunciation variants of European city names was presented (Schaden, 2002). It includes names from a number of European languages (English, German, French, Italian, and Dutch) that were pronounced by speakers from different native language backgrounds, and thus represents potential pronunciation variants of numerous native language/target language (L1/L2) combinations. This speech data collection was compiled in the framework of an ongoing research project dealing with non-native pronunciation variants of European city names. It is one of the aims of this project not only to identify the most common pronunciation errors occurring in various language directions (i.e. L1/L2 pairs), but also to model these variants on the lexicon level by applying phonological rule sets that systematically introduce selected pronunciation errors into canonical lexicons (Schaden, 2003). Rule sets of this type have been compiled for various language directions and are constantly being updated and improved.
The present contribution exhibits some results of this approach by introducing automatically generated accent lexicons of town names for various language directions. The languages presently included are English, German, and French in different L1/L2 combinations. For each of these languages, sample lexicons of 1.000 town names were compiled on the basis of random selections from the GEOnet place names database (http://www.nima.mil/gns/html). The names were phonetically transcribed according to the canonical pronunciation of the corresponding languages and finally converted into potential non-native pronunciation lexicons, where for each native speaker group a separate lexicon was generated. Based on the concept of accent levels introduced in Schaden (2003), multiple pronunications that reflect varying L2 proficiency levels of L1 speakers are derived for each lexicon entry. Currently, four prototypical accent levels plus the canonical form are distinguished; hence each lexicon contains 5.000 pronunciations.
Although similar research (on a considerably larger scale than in the present study) has already been carried out within the framework of the European ONOMASTICA project (viz. the "Onomastica Interlanguage Pronunciation Lexicon"; cf. Onomastica Consortium, 1995), the data compiled in the latter project has unfortunately never been made available to the research community. The CrossTowns sample lexicons will be accessible online (see link on top) in order to stimulate feedback on the quality and validity of the transcriptions. Researchers or any interested persons - particularly speakers of the native languages involved in the study - are invited to download the material and to comment on the proposed pronunciations, to suggest improvements, point out implausible or unlikely variants, and to add or submit important variants or typical mispronunciations they consider to be missing in the transcriptions. This kind of feedback is expected to be a valuable contribution to the improvement of the underlying rule sets that generated the lexicons.
In the paper, the procedures involved in the data collection, an outline of the rule-based accent generation technique, a discussion of the problems involved in modelling non-native pronunciations on the lexicon level, as well as an outlook on some possible ways to evaluate the approach will be presented.
Dahlbäck, N.; Swamy, S.; Nass, C.; Arvidsson, F.; and Skågeby, J., 2001. "Spoken Interaction with Computers in a Native or Non-native Language - Same or Different?" Proceedings INTERACT 2001.
Onomastica Consortium (1995): "The Onomastica Interlanguage Pronunciation Lexicon". Proceedings Eurospeech 1995, Madrid, Spain, 829-832.
Schaden, S. (2002): "A database for the analysis of cross-lingual pronunciation variants of European city names". Proceedings Third International Conference on Language Resources and Evaluation (LREC 2002), Las Palmas de Gran Canaria, Spain, Vol. 4, 1277-1283.
Schaden, S. (2003): "Generating Non-Native Pronunciation Lexicons by Phonological Rules". Proceedings 15th International Conference of Phonetic Sciences (ICPhS 2003), Barcelona, Spain.
van Compernolle, D. (1999): "Speech Recognition by Goats, Wolves, Sheep and Non-natives." Proceedings Workshop on Interoperability in Speech Technology. Leusden, The Netherlands.