Large language models are increasingly used to post-process automatic speech recognition (ASR) output. According to a new arXiv abstract, these models can leverage linguistic context to compare N-best hypotheses and rewrite the transcript when necessary, often reducing word error rate without retraining the recognizer itself.

The same abstract notes that prompt-only correction has limitations, though the sentence is cut off in the source. To address this, the paper proposes "Select or Rewrite," an oracle-guided fine-tuning strategy. The name suggests the model is trained to decide between selecting a hypothesis from the N-best list and generating a rewrite, guided by an oracle signal during fine-tuning.

Because the source is only a partial abstract, details on the oracle construction, training data, and experimental results are not available here. The contribution, as stated, is a fine-tuning framework that explicitly teaches the model when to keep an existing hypothesis and when to rewrite, rather than relying on prompting alone.