This paper presents a study on the automatic characterization of negative emotion in a context of voice service, based on the calculation of acoustic parameters from the speech signal and a LDA classifier. The novelty of this study lies in the application-specific constraints (type of dialog, ASR connection) imposed on the emotion characterization system, and in the solutions proposed to this problem (notably by taking into account the speaker gender and the type of word). Promising results are achieved with acoustic models that are simultaneously gender-specific and words-specific.