Search papers, labs, and topics across Lattice.
1
0
4
2
LLM-based ASR can be shrunk to 2.3B parameters and still beat larger models in real-world scenarios by carefully delineating encoder and LLM roles and using a multi-stage training approach.
