Aligning Very Long Speech Signals to Bilingual Transcriptions of Parliamentary Sessions

Abstract

In this paper, we describe and analyse the performance of a simple approach to the alignment of very long speech signals to acoustically inaccurate transcriptions, even when two different languages are employed. The alignment algorithm operates on two phonetic sequences, the first one automatically extracted from the speech signal by means of a phone decoder, and the second one obtained from the reference text by means of a multilingual grapheme-to-phoneme transcriber. The proposed algorithm is compared to a widely known state-of-the-art alignment procedure based on word-level speech recognition. We present alignment accuracy results on two different datasets: (1) the 1997 English Hub4 database; and (2) a set of bilingual (Basque/Spanish) parliamentary sessions. In experiments on the Hub4 dataset, the proposed approach provided only slightly worse alignments than those reported for the state-of-the-art alignment procedure, but at a much lower computational cost and requiring much fewer resources. Moreover, if the resource to be aligned includes speech in two or more languages and speakers conmute between them at any time, applying a speech recognizer becomes unfeasible in practice, whereas our approach can be still applied with very competitive performance at no additional cost.

This work has been supported by the University of the Basque Country, under grant GIU10/18 and project US11/06, by the Government of the Basque Country, under program SAIOTEK (project S-PE11UN065), and the Spanish MICINN, under Plan Nacional de I+D+i (project TIN2009-07446, partially financed by FEDER funds).