Tunable Q wavelet transform -based features in the classification of phonation types in the singing and speaking voice

Phonation is the use of the laryngeal system, with the help of an air-stream provided by the respiratory system, to generate audible sounds. Humans are capable of generating voices of various phonation types (e.g., breathy, neutral and pressed), and these types are used both in singing and speaking. In this study, we propose to use features derived using the tunable Q wavelet transform (TQWT) for classification of phonation types in the singing and speaking voice. In the proposed approach, the input voice signal is first decomposed into sub-bands using TQWT, and then the Shannon wavelet entropy of each sub-band is calculated. A Feed forward neural network (FFNN) classifier is trained using the entropy values to discriminate three phonation types (breathy, neutral and pressed). The results show that the proposed TQWT-based features outperformed six state-of-the-art features in classification of phonation types both in the singing and speaking voice. Furthermore, the TQWT features achieved the highest phonation classification accuracies of 91% and 82% for the singing and speaking voice, respectively.