The similarity metric employed (Kondrak, 2002) describes segments in terms of Ladefoged's (1995) multivalued features1. This is a feature system that reflects the continuous nature of articulatory differences between segments. In the similarity metric, features are weighted by salience, which is a measure of the features' impact on similarity.
Kondrak provides features specifications for 26 phonemes covering common phonemes in the world's languages. The phoneme inventory as well as their features and feature values can be adapted to match more closely the language under study.
We created feature specifications for the phonemes in our Spanish, English and Basque audio alignment systems.
The salience weights used for all three languages were:
1The subset of Ladefoged features used by Kondrak (2002) is described in the thesis, p. 56. For more information on Ladefoged's feature system, an option is this freely available e-book-chapter by Dan Mateescu (English Phonetics and Phonological Theory).