Enhancing Multilingual Voice Toxicity Detection with Speech-Text Alignment