SAVAS · Sharing AudioVisual language resources for Automatic Subtitling

Due to recently approved European and National directives and laws, the subtitling demand has grown fast in the past few years throughout Europe. The path of manual subtitling is no longer feasible, due to the quantity of the demand and the cost of the process, both in terms of time and personnel. As a result, broadcasters and subtitling companies are seeking for subtitling alternatives more productive than the traditional manual process. Large Vocabulary Continuous Speech Recognition (LVCSR) is proving to be a useful technology for such a purpose. Respeaking – a technique in which a professional listens to the source audio and re-speaks it to a speech recognition engine which transcribes it – is consolidating as the main subtitling technique employed for live and pre-recorded broadcast productions. Another trend in use today is the application of speech recognition to automatically generate a transcript of a programme's soundtrack without the need of a respeaker, and to use this as the basis of subtitles. Unfortunately, the high cost associated to the collection and annotation of the speech and text corpora required to train each LVCSR system for respeaking and/or automatic transcription has hindered the development of new languages and application domains. However, in order to comply with the new audiovisual legal framework, European broadcasters and subtitling companies are generating speech and text corpora suitable for LVCSR developments on a daily basis. SAVAS aims to acquire, share and reuse audiovisual resources of broadcasters and subtitling companies so that high-tech European ASR companies can use the shared data to develop domain specific LVCSRs and/or LVCSRs in new languages to solve the automated subtitling needs of the media industry. Within the project, data and LVCSR technology for automated subtitling will be collected, shared and developed for the following six languages: Basque, Spanish, Italian, French, German and Portuguese.