Multitask Learning for Grapheme-to-Phoneme Conversion of Anglicisms in German Speech Recognition

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3242–3249 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3242 Multitask Learning for Grapheme-to-Phoneme Conversion of Anglicisms in German Speech Recognition Julia Pritzen1,2, Michael Gref 1, Dietlind Z¨uhlke2, Christoph Schmidt1 1Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS), Germany 2TH K¨oln - University of Applied Sciences, Cologne, Germany {julia.pritzen, michael.gref, christoph.andreas.schmidt}@iais.fraunhofer.de dietlind.zuehlke@th-koeln.de Abstract Anglicisms are a challenge in German speech recognition. Due to their irregular pronunciation compared to native German words, automatically generated pronunciation dictionaries often contain incorrect phoneme sequences for Anglicisms. In this work, we propose a multitask sequence-to-sequence approach for grapheme-to-phoneme conversion to improve the phonetization of Anglicisms. We extended a grapheme-to-phoneme model with a classification task to distinguish Anglicisms from native German words. With this approach, the model learns to generate different pronunciations depending on the classification result. We used our model to create supplementary Anglicism pronunciation dictionaries to be added to an existing German speech recognition model. Tested on a special Anglicism evaluation set, we improved the recognition of Anglicisms compared to a baseline model, reducing the word error rate by a relative 1 % and the Anglicism error rate by a relative 3 %. With our experiment, we show that multitask learning can help solving the challenge of Anglicisms in German speech recognition. Keywords: Automatic speech recognition, grapheme-to-phoneme models, sequence-to-sequence models, multitask learning, Anglicisms 1. Introduction Pronunciation dictionaries of hybrid automatic speech recognition (ASR) and Text-to-Speech (TTS) systems, which contain all word forms, are usually generated automatically using a grapheme-to-phoneme (G2P) model trained on a baseline dictionary. Being trained on one language only, monolingual G2P models of- ten struggle with loanwords of foreign heritage since their pronunciation is derived from their source lan- guage. In the German language, the use of English loanwords (Anglicisms) has been steadily increasing (Burmasova, 2010), resulting in a ratio of 4.5 % An- glicisms in spontaneous speech (Hunt, 2019). German G2P models show higher error rates for Anglicisms due to their irregular pronunciation compared to native Ger- man words (Milde et al., 2017). Currently, much of the ASR research has moved away from hybrid ASR systems and more towards end-to- end modeling. However, for real-world applications in non-English languages, we believe there is still a need for improved phonetic pronunciation. In end-to-end approaches, adaptation to the non-canonical pronun- ciation of words, such as Anglicisms, would mainly be achievable via fine-tuning. This requires sufficient amounts of transcribed speech with Anglicisms, which, in general, is not always available. The pronuncia- tion lexicons for training G2P systems are usually more common and easier to obtain. Generating phoneme sequences with a G2P model trained on German data for Anglicisms can lead to wrong pronunciations. Looking at the Anglicism “Whistleblower”, a conventional Sequitur G2P system trained on PHONOLEX Core1 (Schiel, 1998) generates the pronunciation /v I s t l e: p l o 6/ by applying the pronunciation rules it learned from the German training data, …

Content truncated to 500 words