Improving Joint Speech-Text Representations Without Alignment