Adaptation and training of a Swiss German speech recognition system using data-driven pronunciation modeling

* Presenting author
Day / Time: 20.03.2018, 15:40-16:00
Room: MW 1050
Typ: Poster
Das Posterforum zu diesen Postern findet am Dienstag von 16:00 - 16:40 Uhr direkt am Poster in der Magistrale des MW-Gebäudes - Posterinsel G - statt. Die Poster hängen von Dienstag morgen bis Donnerstag nachmittag.
Abstract: Automatic speech recognition is a very important technique for numerous applications like automatic subtitling, dialogue systems and information retrieval systems. Given an annotated speech corpus, a phonetic lexicon and a text corpus, the training of speech recognition systems is straight forward. However, sometimes some of these resources are not available, and strategies must be explored to fill this gap. In this work we train a Swiss German speech recognition system. The only resources that are available is a small Swiss German speech corpus, which is annotated with standard German text. Standard German is the desired output of the speech recognition system, since there is no standardized way to write Swiss German. We use a data-driven approach to estimate the Swiss German pronunciations from a standard German speech recognition model, to improve the Swiss German speech recognition system. Evaluations of the Swiss German speech recognition system show promising results.