LaSR: Context-Aware Speech Recognition via Latent Reasoning

In three linesLaSR proposes a training paradigm for Speech LLMs featuring latent reasoning aligned to acoustic feature regions of target words. Without explicit intermediate tokens, the method improves specialized vocabulary recognition on Fun-Audio-Chat. A new Spoken Darwin-Science corpus for academic terminology is introduced.

Summary generated by Claude — human-verified