Third International Conference on Spoken Language Processing (ICSLP 94)
The results of a language identification experiment employing speech material covering four European languages - Danish, English, German and Italian - are presented and discussed.
The methodology applied presupposes that the languages under test have some speech sounds which are similar enough to be equated across these languages. These speech sounds are the cross-language polyphonemes. Contrasting them there exists a set of non-combinable language dependent monophonemes for each of the languages. The poly- and monophonemes are separated on the basis of a combined database covering training speech material from all four languages. In the language identification experiment each of the four languages is modelled by a set of CDHMMs covering the combined set of language-specific monophonemes and the cross-language polyphonemes. The results show an average language identification score of 88.1% covering the range from 77.9% for the lowest to 95.5% for the highest identification score.
Bibliographic reference. Dalsgaard, Paul / Andersen, Ove (1994): "Application of inter-language phoneme similarities for language identification", In ICSLP-1994, 1903-1906.