Modeling Dialectal Variation for Swiss German Automatic Speech Recognition

We describe a speech recognition system for Swiss German, a dialectal spoken language in German-speaking Switzerland. Swiss German has no standard orthography, with a significant variation in its written form. To alleviate the uncertainty associated with this variability, we automatically generate a lexicon from which multiple written forms of a given word in any dialect can be generated. The lexicon is built from a small (incomplete) handcrafted lexicon designed by linguistic experts and contains forms of common words in various Swiss German dialects. We exploit the powerful speech representation of self-supervised acoustic pre-training (wav2vec) to address the low-resource nature of the spoken dialects. The proposed approach results in an overall relative improvement of 9% word error rate compared to one based on an expert-generated lexicon for our TV Box voice assistant application.