Prosody modeling has garnered significant attention from the speech processing community. Recent developments in multilingual latent spaces for representing linguistic and acoustic information have become a new trend in various research directions. Therefore, we decided to evaluate the ability of multilingual acoustic neural embeddings and knowledge-based features to preserve sentence-mode-related information at the suprasegmental level. For linguistic information modeling, we selected neural embeddings based on word- and phoneme-level latent space representations. The experimental study was conducted using Italian, French, and German audiobook recordings, as well as emotional speech samples from EMO-DB. Both intra- and inter-language experimental protocols were used to assess classification performance for uni- and multimodal (early fusion approach) features. For comparison, we used a sentence mode prediction system built on top of automatically generated WHISPER-based transcripts.