Fact · Speech recognition
12times faster than real time
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.
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“Dataset NLE AR CTC ... Speed Metrics RTFx: Batch size 96 1722 430 2584 RTFx: Batch size 1 322 12 760”
NLE: Non-autoregressive LLM-based ASR by Transcript Editing, Avihu Dekel et al., arXiv, 2026, Section 5.1, Table 1.
Open the original (DOI 10.48550/arXiv.2603.08397)- What it measures
- The paper states that all RTFx measurements use offline batched inference on a single H100 GPU with bf16 precision. RTFx is the audio duration divided by the processing time, so higher is faster.
- What it does not show
- A preprint from one lab's own measurements. The speed depends on batch size, so the batch size is part of each label, and batch 1 is the single-utterance case the paper emphasizes.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as Dekel et al. 2026, AR batch 1: 12 times faster than real time.
- System
- Controlled AR baseline (batch 1)
- Kind
- Research system
- Hardware
- NVIDIA H100 GPU
- Test set
- Open ASR Leaderboard (English average)
- Year
- 2026