Fact · Voice and AI
Feeding a language model word confusion networks from the speech recognizer's lattice, instead of only its top transcript, helped bridge the gap to an oracle upper bound in spoken question answering and intent classification.
Checked against its source on October 17, 2026
In the source's own words
“bridging the SLU performance gap between using the top ASR hypothesis and an oracle upper bound”
Towards ASR Robust Spoken Language Understanding Through In-Context Learning With Word Confusion Networks, Kevin Everson et al., arXiv preprint 2401.02921, 2024, Abstract.
Open the original- What it measures
- A word confusion network keeps the words the recognizer considered alongside the top guess. The oracle upper bound is the score a model would get with the correct transcript.
- What it does not show
- This is an in-context learning experiment on specific benchmarks, and the gap closed is only part of the total. It shows that the recognizer's uncertainty can help a model, not that a typical chat app uses it.