We address noise-robust "auditory scene understanding" for a robot defined by extracting 6W (What, When, Where, Who, Why, hoW) information on the surrounding environment. Although such a robot has been studied in the field of robot audition, only the first four Ws except for "why" and "how" were in scope. Thus, this paper mainly focuses on extracting "how" information, in particular, on cooking scenes to realize a cooking support robot. In this case, "how" information is regarded as a cooking procedure, we construct sound-based cooking procedure recognition based on two models. One is a conventional statistical model, Gaussian Mixture Model (GMM), which is used for an acoustic model to recognize a cooking sound event such as stirring, cutting and so on. The other is a Hierarchical Hidden Markov Model (HHMM), which is used for a recipe model to recognize a sequence of cooking events, i.e., a cooking procedure. We constructed a prototype system for cooking recipe and procedure recognition. Preliminary results showed that the proposed GMM-HHMM based system outperformed a conventional GMM-HMM based system in terms of noise-robustness in cooking recipe recognition and our system was able to correct misrecognition of cooking sound events using recipe model in cooking procedure recognition.