Bilingual prosodic dataset compilation for spoken language translation

Abstract

This paper builds on a previous methodology that exploits dubbed media material to build prosodically annotated bilingual corpora. The almost fully-automatized process serves for building data for training spoken language models without the need for designing and recording bilingual data. The methodology is put into use by compiling an English-Spanish parallel corpus using a recent TV series. The collected corpus contains 7000 parallel utterances totaling to about 10 hours of data annotated with speaker information, word-alignments and word-level acoustic features. Both the extraction scripts and the dataset are distributed open-source for research purposes.