Excessive Daytime Sleepiness (EDS) is associated with several diseases and therefore negatively affects the daily life of impacted people. Its diagnosis and follow-up are difficult because they require testing at the hospital for one full day. Monitoring patients regularly in ecological conditions may be done through speech analysis. Although several corpora containing speech from sleepy subjects exist, they do not suit ecological requirements regarding either the device used for recording or the speech elicitation tasks. In this paper, we introduce the Medispeech corpus containing reading, daily-life semi-spontaneous, and medically-oriented spontaneous tasks. Fifty-nine French subjects were recorded with both a professional-quality microphone and a smartphone using a dedicated application, resulting in 1,729 recordings for a total duration of 21 hours. Their EDS diagnosis was assessed by both a physiological objective measurement (mean sleep latency measured during a clinical test) and a subjective questionnaire (Karolinska Sleepiness Scale). Phenotyping of subjects is assured by collecting socio-demographic and medical data related to diverse dimensions of sleepiness, comorbidities, and addictions. Finally, we analyse the validity of our data collection protocol by measuring the effective duration of speech (after discarding pauses) and assessing its links with the collected subjects’ characteristics.

Large Language Models (LLMs) have shown remarkable performance in Natural Language Processing tasks, including Machine Translation (MT). In this work, we propose a novel MT pipeline that integrates emotion information extracted from a Speech Emotion Recognition (SER) model into LLMs to enhance translation quality. We first fine-tune five existing LLMs on the Libri-trans dataset and select the most performant model. Subsequently, we augment LLM prompts with different dimensional emotions and train the selected LLM under these different configurations. Our experiments reveal that integrating emotion information, especially arousal, into LLM prompts leads to notable improvements in translation quality.

La transcription phonémique automatique de la parole spontanée trouve des applications variées, notamment dans l’éducation et la surveillance de la santé. Ces transcriptions sont habituellement évaluées soit par la précision de l’identification des phonèmes, soit par leur segmentation temporelle. Jusqu’à présent, aucun système n’a été évalué simultanément sur ces deux tâches. Cet article présente l’évaluation d’un système de transcription phonétique du français spontané (corpus Rhapsodie) basé sur Kaldi. Ce système montre de bons résultats en identification des phonèmes et de leurs catégories, avec des taux d’erreur de 19,2 et 13,4 respectivement. Il est cependant moins performant en segmentation, manquant en moyenne 40 de la durée des phonèmes et 34 des catégories. Les performances s’améliorent avec le niveau de planification de la parole. Ces résultats soulignent le besoin de systèmes de transcription phonétique automatique fiables, nécessaires à des analyses plus approfondies de la parole spontanée.