sStoryCloze

The sStoryCloze refers to the Spoken StoryCloze benchmark, which is a spoken version of the StoryCloze dataset. The StoryCloze dataset consists of five-sentence commonsense stories, where the task is to predict the ending of a story given the first four sentences and two possible endings, one of which is the correct ending and the other is a distractor. The sStoryCloze evaluates the model's capabilities to capture fine-grained causal and temporal commonsense relations in spoken language. It assesses the model's ability to generate coherent and contextually appropriate continuations given a spoken prompt. The dataset is used to evaluate the performance of SpeechLMs in understanding and generating spoken narratives.