Prediction of Creaky Speech by Recurrent Neural Networks Using Psychoacoustic Roughness

Abstract

The use of a psychoacoustic roughness model as a predictor of creaky voice is reported. We found that the roughness temporal profile of vocalic segments can predict the presence of creakiness in speech. Using a simple bi-directional Recurrent Neural Network ( rnn ), we were able to predict the presence of creakiness in vocalic segments from only roughness traces with an accuracy similar to that obtained with rnn s trained on at least 12-dimensional input data (including amplitude difference between the first two harmonics, residual peak prominence, etc.). Training rnn s with the combination of roughness and multidimensional input data improved the performance of the predictor, but not significantly. Likewise, augmenting the dataset by time derivatives of the input features did not improve the predictor's performance. The proposed roughness-based predictor eases interpretation and comparison of creakiness among corpora and suggests that roughness prediction models could be successfully used for classification of creaky intervals in speech.