Combining Multilingual Resources and Models to Develop State-of-the-Art E2E ASR for Swedish

In terms of automatic speech recognition (ASR), Swedish belongs to the group of less-resourced languages, as publicly available training data is limited to a few hundred hours of mostly read speech. To acquire larger amounts of more realistic data, we investigate the existing multilingual approaches, and also propose two new ones, which combine Swedish with previously created Norwegian data and models. We use them for efficient automatic harvesting of spoken Swedish from broadcast, parliament, YouTube, and audiobook archives. The combined models significantly speed up the harvesting process and improve the final Swedish end-to-end (E2E) ASR system. We evaluate it on datasets covering various applications and domains; they provide performance better than the state-of-the-art commercial cloud services. We have made all of our test datasets publicly available for future comparative experiments.