Direct Speech to Speech Turkish to Arabic

Abstract

Dubbed series are gaining a lot of popularity in recent years with strong sup-port from major media services providers. Such popularity is fueled by studiesthat showed that dubbed versions of TV shows are more popular than theirsubtitled equivalents. In this paper, we propose an unsupervised approach toconstruct speech-to-speech corpus, aligned on short segment level, to produce aparallel speech corpus in the source- and target- languages. Our methodologyexploits speech recognition, machine translation and noisy frames removal algo-rithms, to match segments in both languages. Without losing any generalization,our approach was successfully applied on Turkish-Arabic dubbed series. Out of36 hours, our pipeline was able to generate 17 hours of paired segments with 70%overall accuracy. The corpus will be freely available for the research community.