The automatic translation of speech into pictogram terms (Speech-to-Pictos) represents a novel NLP task with the potential to enhance communication for individuals with language impairments. Recent research has not explored the adaptation of state-of-the-art methods to this task, despite its significance. In this work, we investigate two approaches: (1) the cascade approach, which combines a speech recognition system with a machine translation system, and (2) the end-to-end approach, which tailors a speech translation system. We compare state-of-the-art architectures trained on an aligned speech-topictogram dataset, specially created and released for this study. We conduct an in-depth automatic and human evaluation to analyze their behavior on pictogram translation. The results highlight the cascade approach's ability to generate relevant translations from everyday read speech, while the end-to-end approach achieves competitive results with challenging acoustic data.