Speech understanding in large language models usually requires bridging continuous audio and discrete text. A new arXiv paper, "Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs," takes up this integration problem. The authors note that pretrained LLMs provide a powerful reasoning backbone, but continuous acoustic signals do not fit naturally into their token-based processing.
At the center of the work is a design question: how should speech be tied to the LLM? The title suggests one answer — using text as a universal bottleneck, meaning all audio is converted to text before the model processes it. The abstract frames this as a central choice, but the full argument and experimental evidence are not visible in the available text. The paper appears to be a contribution to an ongoing debate about how much acoustic information an LLM truly needs.