Peer-reviewed research
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models (full text, HTML)
Chen Chen et al., arXiv (NeurIPS 2023 Datasets and Benchmarks track), 2023
Read the originalRead on October 17, 2026
- What it is
- The full text of the HyPoradise paper, which builds a benchmark of recognizer n-best lists paired with reference transcripts so that large language models can be trained to correct speech recognition output. Its results tables report relative word error rate reductions on several test sets.
- How much weight it carries
- Peer-reviewed and accepted at NeurIPS 2023. This HTML copy is the arXiv version 1, and the paper's dataset size differs between arXiv and the conference version, so check numbers against the version you need. The reductions are relative, on a few specific test sets (ATIS, WSJ, CHiME-4 and accented speech), and were measured with a 2023 recognizer.
Facts taken from this source
- In the HyPoradise benchmark, the H2T-LoRA language model correction reduced word error rate by 77.1 percent relative on the ATIS test set and by 55.1 percent relative on WSJ.
“H2T-LoRA respectively reduces 77.1% and 55.1% relative WER on ATIS and WSJ” (Results, Table 2)