Speech Translation with Style: AppTek’s Submissions to the IWSLT Subtitling and Formality Tracks in 2023

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 251–260 July 13-14, 2023 c⃝2023 Association for Computational Linguistics Speech Translation with Style: AppTek’s Submissions to the IWSLT Subtitling and Formality Tracks in 2023 Parnia Bahar∗, Patrick Wilken∗, Javier Iranzo-Sánchez, Mattia di Gangi, Evgeny Matusov, Zoltán Tüske Applications Technology (AppTek), Aachen, Germany {pbahar,pwilken,jiranzo,mdigangi,ematusov,ztuske}@apptek.com Abstract AppTek participated in the subtitling and for- mality tracks of the IWSLT 2023 evaluation. This paper describes the details of our subti- tling pipeline - speech segmentation, speech recognition, punctuation prediction and inverse text normalization, text machine translation and direct speech-to-text translation, intelligent line segmentation - and how we make use of the provided subtitling-specific data in training and fine-tuning. The evaluation results show that our final submissions are competitive, in par- ticular outperforming the submissions by other participants by 5% absolute as measured by the SUBER subtitle quality metric. For the for- mality track, we participated with our En-Ru and En-Pt production models, which support formality control via prefix tokens. Except for informal Portuguese, we achieved near perfect formality level accuracy while at the same time offering high general translation quality. 1 Introduction This paper presents AppTek’s submissions to the subtitling and formality tracks of the IWSLT 2023 evaluation campaign. In the subtitling track, we participate in constrained and unconstrained condi- tions and in both language pairs English-to-German (En-De) and English-to-Spanish (En-Es). In the formality track, we participate in the zero-shot un- constrained condition for English-to-Portuguese (En-Pt) and English-to-Russian (En-Ru). This paper is organized as follows: Section 2 briefly describes our data preparation. Section 3 presents AppTek’s pipeline for subtitle translation. Its different components, namely audio segmen- tation, speech translation (ST), automatic speech recognition (ASR), machine translation (MT) mod- els, and our subtitle segmentation algorithm are described in Sections 3.1-3.5. Section 3.6 contains experiments and an analysis of our subtitling sys- tems. Section 4 presents AppTek’s approach to ∗equal contribution formality-controlled machine translation. Finally, Section 4.1 shows the results of our formality track submission. 2 Data Preparation 2.1 Text Data We use all of the allowed “speech-to-text paral- lel” and “text-parallel” data, including Europarl, Europarl-ST, News Commentary, CORDIS News, Tatoeba, TED2020, IWSLT TED, MuST-C v3, CoVoST v2, and OpenSubtitles1. We apply com- mon parallel data filtering steps based on lan- guage identification, sentence length ratios between source and target sentences and additional heuris- tics. After filtering, we obtain 13.5M sentence pairs with 152M running words (counted on the English side) for En-De and 16.5M sentence pairs with 183M words for En-Es. Next, we clone this data and process the En side of the clone with our text normalization tool NEWTN. It implements elaborate regular expres- sions to convert numbers, dates, monetary amounts, and other entities with digits into their spoken form. It is also used to remove punctuation and word case information. After training on such source data, our MT systems are able to directly translate from raw ASR output that lacks punctuation and casing into properly formatted written target language text. …

Content truncated to 500 words