Dr. Christina Esposito

Although several measures of phonation have been demonstrated to accurately predict phonation in certain sets of languages (e.g. H1H2), there is no known system by which phonation types can be reliably classified cross-linguistically. These researcher employed a large genetic-algorithm-based ensemble classifier trained on a large multilingual speech database (the UCLA Production and Perception of Voice Quality dataset) to identify the language spoken in a given audio segment. The feature representations used in the language-identification model are low-dimensional, fuzzy representations of the acoustic signal, so the features of the held-out language can be approximated by comparing the audio signal to the learned representations without training a new classifier. The resultant feature representations of each language are then used in a similar ensemble classifier system to aid in the classification of phonation in a held-out language.