Preprint
Towards ASR Robust Spoken Language Understanding Through In-Context Learning With Word Confusion Networks
Kevin Everson et al., arXiv preprint 2401.02921, 2024
Read the originalRead on October 17, 2026
- What it is
- A study of giving a language model more than the single best speech recognition result. The authors feed it word confusion networks built from the recognizer's lattice, which keep the alternative words the recognizer considered, and test spoken question answering and intent classification.
- How much weight it carries
- A preprint from a large speech research group, with an extended team of authors. It compares against an oracle upper bound and reports the gap, but the figures are in the full paper and are not repeated here. Results are for in-context learning on specific benchmarks.
Facts taken from this source
- 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.
“bridging the SLU performance gap between using the top ASR hypothesis and an oracle upper bound” (Abstract)