Measuring Norwegian dialect distances using acoustic features

Computational dialectometry has been proven to be useful for finding dialect relationships and identifying dialect areas. The first to develop a method of measuring dialect distances was Jean Séguy, assisted and inspired by Henri Guiter (Chambers and Trudgill, 1998). Strongly related to the methodology of Séguy is the work of Goebl, although the basis of Goebl’s work was developed mainly in dependently of Séguy (Goebl, 1982, 1993). In 1995 Kessler used the Levenshtein distance for finding linguistic distances between Irish Gealic dialects, and in 1996 the same algorithm was applied to Dutch dialects by Nerbonne et al.. The The Levenshtein distance is a sensitive measure with which distances between strings (in this case transcriptions of word pronunciations) are calculated. Gooskens and Heeringa (2004) showed that linguistic dialect distances measured with Levenshtein correlate significantly with perceptual distances for 15 Norwegian varieties (r = 0.67, p < 0.001). Pronunciation based dialect distance measurements used in previous studies are based on phonetic transcriptions. However it is time-consuming to make phonetic transcriptions and furthermore the quality of the transcriptions sometimes varies greatly, depending on the skills of the transcriber. When several transcribers are involved, the data may not be consistent. What the result of this can be, is, for example, shown by Heeringa (2005) who found the Frisian dialect area to be divided in a northern and southern part, which reflects the work areas of the two transcribers. In the field of Automatic Speech Recognition methods can be found with which pronunciations are compared on the basis of the acoustic signal, without intervention of a transcriber. Among others we mention Hunt et al. (1999) and ∗University of Groningen, Department of Information Science †UC Berkeley, Department of Linguistics ‡University of Groningen, Scandinavian Department UC Berkeley Phonology Lab Annual Report (2005)