Comparison of Phonetic Segmentation Tools for European Portuguese

Abstract

Currently, the majority of the text-to-speech synthesis systems that provide the most natural output are based on the selection and concatenation of variable size speech units chosen from an inventory of recordings. There are many different approaches to perform automatic speech segmentation. The most used are based on (Hidden Markov Models) HMM [1,2,3] or Artificial Neural Networks (ANN) [4], though Dynamic Time Warping (DTW) [3,4,5] based algorithms are also popular. Techniques involving speaker adaptation of acoustic models are usually more precise, but demand larger amounts of training data, which is not always available.

In this work we compare several phonetic segmentation tools, based in different technologies, and study the transition types where each segmentation tool achieves better results. To evaluate the segmentation tools we chose the criterion of the number of phonetic transitions (phone borders) with an error below 20ms when compared to the manual segmentation. This value is of common use in the literature [6] as a majorant of a phone error. Afterwards, we combine the individual segmentation tools, taking advantage of their differentiate behavior accordingly to the phonetic transition type. This approach improves the results obtained with any standalone tool used by itself. Since the goal of this work is the evaluation of fully automatic tools, we did not use any manual segmentation data to train models. The only manual information used during this study was the phonetic sequence.