This paper proposes a method based on epoch parameters for detection of creaky voice in speech signal. The epoch parameters characterizing the source of excitation considered in this work are number of epochs in a frame, strength of excitation of epochs and epoch intervals. Analysis of epoch parameters estimated from zero-frequency filtering method with different window sizes is carried out. Distinct variations in the epoch parameters are observed for modal and creaky voiced regions. Variances of epoch parameters are used as input features to train a neural network classifier for identifying creaky regions. Performance evaluation results indicate that the proposed method performs significantly better than the existing creaky detection methods on different speech databases.
Bibliographic reference. Narendra, N. P. / Rao, K. Sreenivasa (2015): "Automatic detection of creaky voice using epoch parameters", In INTERSPEECH-2015, 2347-2351.