This paper deals with the analysis of the emotional content of voice messages left on an answering machine, in the context of a customer care service. The 103 messages which constitute the corpus are split into 478 emotional phases. These phases are then used as material for a subjective test and a signal analysis. The relation between subjective data and acoustical data is then explored with a focus drawn on inter-subject agreements obtained in the subjective test and on speakers' gender. The results show that the acoustical parameters which are relevant to differentiate emotions for male speakers differ from those able to differentiate emotions for female speakers, but also depend of the inter-subject agreement, which governs the number of observations.