We have developed a framework to record spontaneous speech of infants. Using the framework, we have accumulated infant speech data, and proved that the data is quite efficient for the explication of the spoken language acquisition process. We aim at constructing the "multimodal infant behavior corpus," which contributes to the elucidation of human commonsense knowledge and its acquisition mechanism. We previously established the environments to record infant behaviors as multimodal data. We have newly developed a wearable speech recording system and succeeded to record infant utterances with high quality. We have held an infant school once a week for 10 months, and accumulated infant speech data more than 100 hours long. We observed infant utterances in the aspects of acoustic and prosodic features. Through an acoustic observation, we have succeeded to analyze alteration of the pronunciation skills focused on demonstratives that appear quite often in infant utterances. As the result of a prosodic observation, we have also obtained knowledge of how infants enrich and diversify the ways to explain their intentions or emotions corresponding to their growth.