This paper proposes a new procedure based on a maximum likelihood approach using hidden Markov models to detect infants’ emotions through their cries. Our procedure uses stochastic acoustic models for each kind of emotion. The acoustic models are generated using infant’s cries that are labeled segmentally according to their acoustic features. The procedure detects segment sequences with the highest likelihood among all kinds of emotions. The results of our preliminary recognition experiments on two emotions using three types of segment labeling show that the proposed procedure is applicable to emotion detection in an infant’s cry and that the detailed transcription of acoustic segments is useful. In this paper, using detailed transcriptions, we broaden the experiments to include five emotions. Assuming the judgment of each infant’s mother to be correct, we compared the result of the experiment and that of a subjective opinion test. We conducted the opinion test with the help of three child-rearing experts. Emotion recognition using the proposed procedure displays a favorable comparison with the judgment of our experts, showing the validity of the proposed procedure.