How well can LLMs handle novel phonetic forms?

Authors: Daniel Duran, Laurens Winkler, Sina Zarrieß

Abstract:

We present an analysis of tools for automatic transcription. Our focus is on the utility of speech-to-text (automatic speech recognition, ASR) and speech-to-phone recognition systems in transcribing spontaneous speech which includes novel phonetic forms. Our results show that the transcription quality of novel phonetic forms is significantly worse in comparison to regular control items.