Matesub: The Translated Subtitling Tool at the IWSLT2023 Subtitling Task

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 461–464 July 13-14, 2023 c⃝2023 Association for Computational Linguistics Matesub: the Translated Subtitling Tool at the IWSLT2023 Subtitling task Simone G. Perone Translated srl via Indonesia 23 00144 Rome - Italy simone@translated.com Abstract This paper briefly describes Matesub, the subti- tling tool Translated used to participate in the Subtitling shared task at IWSLT 2023. Matesub is a professional web-based tool that combines state-of-the-art AI with a WYSIWYG editor. The automatic generation of subtitles in Mate- sub is based on a cascade architecture, com- posed of ASR, text segmenter and MT neural models, which allows covering any pair from about 60 languages and their variants. 1 Matesub Matesub1 is a web-based tool released by Trans- lated2 that combines state-of-the-art AI with a WYSIWYG (What You See Is What You Get) ed- itor for supporting professionals in the creation of subtitles for audio visual documents. Matesub generates subtitling suggestions through a process- ing pipeline which was used to participate in the Subtitling shared task at IWSLT 2023. This paper first describes the pipeline, and then presents and discusses the scores of the submission. 1.1 The subtitling pipeline In Matesub, subtitles are automatically generated by a pipeline (Figure 1) which concatenates two main modules, based on neural models: an au- tomatic speech recognition (ASR) system and a module providing the Captions & Subtitles Service. They are described in the following. Figure 1: Architecture of the subtitling pipeline. 1https://matesub.com/ 2https://translated.com/ 1.1.1 Automatic speech recognition The ASR is in charge of the transcription of the speech content of an audio signal. In Matesub, this processing stage is provided either by an in- house ASR model or by a 3rd party commercial ASR service, according to the availability of the internal solution and its relative quality. In both cases, the word hypotheses are expected to be given in conversation time mark (CTM) format. This text file consists of records each having 5 fields, e.g.: 23.66 0.29 human 0.00998 False 23.96 0.40 beings. 0.01000 True 24.48 0.13 We 0.33000 False whose meaning is given in Table 1. field meaning 1 start time (sec) 2 duration (sec) 3 token (i.e. word) 4 confidence 5 end of sentence (boolean) Table 1: Fields in the CTM format. Note that the transcription is punctuated and cased; moreover, the flag indicating the end of sen- tence is typically set on for acoustic reasons, like the presence of the pause between the tokens begin. and We, but - less frequently - also for “linguistic” evidence (learned by the ASR from training data). 1.1.2 Captioning and subtitling The Captions and Subtitles Service is in charge of building, starting from a given CTM file, the SubRip Subtitle (SRT) files of the transcription contained in the CTM file and its translation; the two SRTs are finally merged in a single JSON file. As shown in Figure 2, this module consists of two main components, a text segmenter and a neural machine translation (NMT) …

Content truncated to 500 words