Towards an Open-Source Dutch Speech Recognition System for the Healthcare Domain

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1032–1039 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1032 Towards an Open-Source Dutch Speech Recognition System for the Healthcare Domain Cristian Tejedor-Garc´ıa∗, Berrie van der Molen†, Henk van den Heuvel∗, Arjan van Hessen‡, Toine Pieters† ∗CLS/CLST, Radboud University, Nijmegen, the Netherlands {cristian.tejedorgarcia, henk.vandenheuvel}@ru.nl †Freudenthal Institute, Utrecht University, Utrecht, the Netherlands {b.j.vandermolen, t.pieters}@uu.nl ‡HMI, University of Twente, Enschede, the Netherlands a.j.vanhessen@utwente.nl Abstract The current largest open-source generic automatic speech recognition (ASR) system for Dutch, Kaldi NL, does not include a domain-specific healthcare jargon in the lexicon. Commercial alternatives (e.g., Google ASR system) are also not suitable for this purpose, not only because of the lexicon issue, but they do not safeguard privacy of sensitive data sufficiently and reliably. These reasons motivate that just a small amount of medical staff employs speech technology in the Netherlands. This paper proposes an innovative ASR training method developed within the Homo Medicinalis (HoMed) project. On the semantic level it specifically targets automatic transcription of doctor-patient consultation recordings with a focus on the use of medicines. In the first stage of HoMed, the Kaldi NL language model (LM) is fine-tuned with lists of Dutch medical terms and transcriptions of Dutch online healthcare news bulletins. Despite the acoustic challenges and linguistic complexity of the domain, we reduced the word error rate (WER) by 5.2%. The proposed method could be employed for ASR domain adaptation to other domains with sensitive and special category data. These promising results allow us to apply this methodology on highly sensitive audiovisual recordings of patient consultations at the Netherlands Institute for Health Services Research (Nivel). Keywords: speech recognition, language modeling, domain adaptation, healthcare 1. Introduction 1.1. Background Every year we encounter more than 15,000 hospital admissions due to avoidable misuse of medicines in the Netherlands (Erasmus MC, Nivel, Radboud UMC, PHARMO, 2017). Often, this has to do with the pa- tient’s unintentional improper use or low levels of ad- herence. Practical limitations and barriers are in the foreground here, over which the patient has insufficient control (van Dijk et al., 2016). For instance, they are not good at reading, understanding (functionally illiter- ate), reproducing (forgetfulness, real cognition), and/or accurately performing the prescribed usage. This might result in rather diverse and inappropriate forms of use, low levels of adherence and waste of scarce financial resources (Maas et al., 2020). In order to overcome these misunderstandings, we need to better understand the explicit and implicit attribution of meaning to medicines as part of the information pro- cessing. For that reason, effective and efficient tran- scriptions of doctor-patient interviews are indispens- able. This, however, is a context with considerable privacy-sensitive constraints. Research on and use of sensitive data involving au- dio/video (AV) recordings requires an infrastructure where both the data and the research environment are optimal in terms of General Data Protection Regulation (GDPR) safeguards. In the HoMed project1 we have set out to establish a method …

Content truncated to 500 words