Semantic enrichment towards efficient speech representations