Search papers, labs, and topics across Lattice.
1
0
3
Speech tokenizers, despite being crucial for multimodal LLMs, primarily capture phonetic information, creating a semantic mismatch with text-derived semantics that hinders performance.
