This thesis concerns the detection of emotions in multi-lingual audio utterances. One application being considered is the indexing of emotional states in audio-visual documents for their search by contents. Our work begins with the study of emotion and of its model representations: discrete, continuous and hybrid models. In the following of the work, only the discrete model will be used for practical reasons linked to evaluation but also because it is easier to use in the targeted applications. A state of the art on the different approaches used for emotion recognition is then presented. The problem of the production of annotated corpus for training and evaluation of emotional state recognition systems is also considered and an overview of the available corpus is given. One of the difficulties on this point is to obtain realistic corpus for the target applications. To obtain data more spontaneous and more diverse in languages, two corpora were created from motion pictures, one in English and one in Vietnamese. The following work is divided into four parts: study and search for the best parameters to represent the acoustic signal for the emotion recognition, study and search for the best models and classification systems for the same problem, experiments on the recognition emotions across languages and, finally, production of an annotated Vietnamese corpus and assessment of emotion recognition in this language which has the specificity of being tonal. In the first two studies, mono-speaker, multi-speaker and speaker-independent cases were considered. The search for the best parameters was performed on a broad set of global and local parameters traditionally used in automatic speech processing as well as derivations them. An approach based on the forward forced sequential selection was used for selecting optimal combinations of acoustic parameters. The same approach can be used on different data types, although the final result depends upon the type. Among the MFCC, LFCC, LPC, fundamental frequency, intensity, phonetic rate and other parameters from the time-domain, MFCC gave the best results in the considered cases. A symbolic normalization approach has helped to improve the performance in the speaker independent case. For the search for the best models and associated classification systems, an approach by successive elimination within cases of increasing complexity (single-speaker, multi-speaker and speaker-independent) was used. The GMM, HMM, SVM and VQ (vector quantization) models have been studied. The GMM model is the one which led to the best results on the considered data. Cross-language experiments (German and Danish) have shown that the developed methods work well from one language to another, but that a specific optimization of the parameters for each language and for each type of data is necessary for obtaining the best results. These languages are not tonal languages, however. Tests with the created Vietnamese corpus have shown a much less good generalization in this case. This may be due to the fact that the Vietnamese language is tonal but it may also be due to the difference between the conditions of creation of the corpora: action in the first case and more spontaneous for the Vietnamese.
