Classification of pathological voices is an important problem for early detection and diagnosis. The automatic analysis is a useful complementary tool to other methods based on direct observation. The pathological voices that are studied in this paper are the voices of patients who have a peripheral facial paralysis which, among other pronunciation impairments, affects their ability to pronounce bilabial sounds. The idea is then to use a burst detector to compute acoustical features and provide them to a classifier in order to automatically determine the degree of the voice impairment. The speech database that is used through the paper is unique and was recorded in an soundproof cabin at the La Pitie Salpetriere Hospital in Paris, France. Even if the database is in French, the features that are used in those experiments are independent from the language. The speech recordings used for the experiments are isolated sentences. Two kinds of artificial neural networks are studied for the classification task, a multilayer perceptron and a neural network based on learning vector quantization. Our results show a correlation between the burst-based acoustic features computed from the voices and the degree of the impairment that affects patients.
Bibliographic reference. Mauclair, Julie / Koenig, Lionel / Robert, Marina / Gatignol, Peggy (2013): "Burst-based features for the classification of pathological voices", In INTERSPEECH-2013, 2167-2171.