Preprint
NLE: Non-autoregressive LLM-based ASR by Transcript Editing
Avihu Dekel et al., arXiv, 2026 · DOI 10.48550/arXiv.2603.08397
Read the originalRead on October 19, 2026
- What it is
- An IBM Research preprint on a non-autoregressive speech recognizer that edits an initial hypothesis in parallel. It reports inverse real-time factors (RTFx) on the Open ASR Leaderboard for its model and for an autoregressive baseline, all measured on a single NVIDIA H100 GPU.
- How much weight it carries
- Preprint, not yet peer-reviewed. The speed measurements are the authors' own, with a stated protocol (offline batched inference, bf16, one H100), and the speed tables are internally consistent with the abstract. Speed depends on batch size, which the paper reports.
Facts taken from this source
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 1, the controlled autoregressive baseline ran at an RTFx of 12, meaning 12 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 96, the controlled autoregressive baseline ran at an RTFx of 430, meaning 430 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 1, the CTC model ran at an RTFx of 760, meaning 760 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 96, the CTC model ran at an RTFx of 2584, meaning 2584 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 1, NLE ran at an RTFx of 322, meaning 322 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 1, NLE++ ran at an RTFx of 310, meaning 310 times faster than real time.
“on a single H100 GPU with bf16 precision ... NLE NLE++ ... RTFx: Batch size 96 1722 1630 RTFx: Batch size 1 322 310” (Section 5.1, Table 2)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 96, NLE++ ran at an RTFx of 1630, meaning 1630 times faster than real time.
“NLE++ achieves 5.67% average WER with an RTFx of 1630. ... on a single H100 GPU with bf16 precision” (Abstract; Section 5.1, Table 2)
- In the paper's Open ASR Leaderboard speed measurements on a single H100 GPU at batch size 96, NLE ran at an RTFx of 1722, meaning 1722 times faster than real time.
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760” (Section 5.1, Table 1)