As conversational search becomes more pervasive, it becomes increasingly important to understand the user's underlying needs when they converse with such systems in diverse contexts. We report on an in-situ experiment to collect conversationally described information needs in a home cooking scenario. A human experimenter acted as the perfect conversational search system. Based on the transcription of the utterances, we present a coding scheme comprising 27 categories to annotate the information needs of users. Moreover, we use these annotations to perform prediction experiments based on random forest classification to establish the feasibility of predicting the information need from the raw utterances. We find that a reasonable accuracy in predicting information need categories is possible.