Discrimination of Sonorants from Fricatives Using a Scalar Feature Derived from Linear Prediction Coefficients

The sum of linear prediction coefficients (LPCs) is proposed as an effective feature in discriminating between sonorants and fricatives in continuous speech. On the closed set of sonorant and fricative frames of the entire TIMIT test database, a classification accuracy of 98.23% is obtained. When this feature is combined with three other features derived from the LPCs, the feature vector achieves an accuracy of 98.27% using a linear support vector machine classifier. The accuracy increases to 98.41% with mel frequency cepstral coefficients also added. The robustness of the feature has been tested on additive white, babble and pink noise.