The PASSAGE project : Standard German Subtitling of Swiss German TV content

The PASSAGE project : Standard German Subtitling of Swiss German TV content Pierrette Bouillon, Johanna Gerlach, Jonathan Mutal, Marianne Starlander Faculty of translation and interpreting, University of Geneva 40, bd du Pont d’Arve, 1211 Geneva, Switzerland {pierrette.bouillon, johanna.gerlach, jonathan.mutal, marianne.starlander}@unige.ch Abstract We present the PASSAGE project, which aims at automatic Standard German subti- tling of Swiss German TV content. This is achieved in a two step process, begin- ning with ASR to produce a normalised transcription, followed by translation into Standard German. We focus on the sec- ond step, for which we explore different approaches and contribute aligned corpora for future research. 1 Introduction Swiss German, a primarily spoken language with many regional dialects and no standardised writ- ten form (Honnet et al., 2018), is spoken by two thirds of the population of Switzerland. It is widely used on Swiss TV, e.g. in news reports or inter- views, which are subtitled in Standard German to make them accessible to people who cannot un- derstand spoken Swiss German. Producing these subtitles automatically would be advantageous in terms of time and cost. This task is the focus of the PASSAGE project (Nov. 2020- Dec. 2022) “Sous-titrage automatique du suisse allemand en allemand standard”, which is a collaboration be- tween Geneva University, SRF (Schweizer Radio und Fernsehen) and recapp.1 In this project a first automatic speech recogni- tion (ASR) step is used to produce a normalised transcription of spoken Swiss German, keeping the original syntax and expressions but using Standard © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1https://www.media-initiative.ch/project/subtitling-of-swiss- german-into-standard-german-automatic-post-editing/ German words. In a second step, different ap- proaches are explored to transform this normalised transcription into correct written Standard German (see Figure 1). To achieve this, multiple issues must be dealt with: ASR errors, incorrect detection of sentence boundaries, features related to spon- taneous spoken language, such as dysfluencies or informal language, and finally the syntactic diver- gences between Swiss German and Standard Ger- man (Glaser and Bart, 2021). The three goals of the project are 1) to create data sets for Swiss Ger- man, 2) to build systems for the translation of ASR output into Standard German, and 3) to evaluate the usability of the system output. 2 Data The following data were provided by SRF: • Normalised transcriptions of TV shows: orig- inally created to train the Swiss German speech recogniser, these human transcriptions keep the original syntax and expressions but use Standard German words. (98,126 seg- ments) • Original Standard German subtitles of the TV shows (DE): batches of subtitles, not aligned with the transcriptions. (101,150 segments) Based on these data, we have so far created sev- eral aligned corpora, which were used to train and specialise the first systems: • Normalised transcriptions - Standard Ger- man: this corpus was produced by manual post-editing of the transcriptions. (20,634 segments) • Normalised transcriptions - original subti- tles: this corpus was aligned automatically. (70,374 segments) Figure 1: Overview of …

Content truncated to 500 words