Massively multilingual and multimodal sentence representations like SONAR are usually trained to capture only the meaning of the encoded text or speech. We complement this semantic embedding by a generic speech characteristic embedding which captures the expressive properties of a speech signal. We describe an iterative training procedure which aims to disentangle the semantics and expressive speech properties, and which does not need labeled data. We show the effectiveness of our method on the FLEURS and mEXPRESSO benchmark test sets using multiple metrics which aim to measure the preservation of the meaning and prosody for zero-shot speech-to-speech translation from five languages into English.