Speech Prosody 2006

Dresden, Germany
May 2-5, 2006

Facing Data Scarcity Using Variable Feature Vector Dimension

Pablo Daniel AgŁero, Antonio Bonafonte

TALP Research Center, Universitat PolitŤcnica de Catalunya (UPC), Barcelona, Spain

This paper focuses on three key points of intonation modelling: interpolation of fundamental frequency contour, sentence by sentence parameter extraction and data scarcity. In some cases, they introduce noise and inconsistency on training data reducing the performance of machine learning techniques.

We consider that the F0 contour is segmented into prosodic units (such as accent groups, minor phrases, etc). Each segment of F0 contour has a corresponding feature vector with linguistic and non-linguistic components.

We propose to face the limitations mentioned above using a technique based on clustering using different feature vector dimensions. The clustering of feature vectors produces also a partition in the F0 contour space. The proposal consists on a procedure to select the dimension that contributes to predict the best fundamental frequency contour from a RMSE sense compared to a reference contour. Experimental results show an improvement compared to other approaches.

Full Paper

Bibliographic reference.  AgŁero, Pablo Daniel / Bonafonte, Antonio (2006): "Facing data scarcity using variable feature vector dimension", In SP-2006, paper 150.