Fact · Speech recognition
46.81times faster than real time
In the MDM-ASR paper, the authors' system ran at an RTFx of 46.81 on LibriSpeech test-clean and test-other, measured on a single NVIDIA A100 GPU at batch size 1, which is 46.81 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
“all models are evaluated on a single NVIDIA A100 GPU with a batch size of 1 ... MDM-ASR (ours) 1.8 3.6 10.7 12.2 6.0 6.9 1.0 46.81”
MDM-ASR: Bridging Accuracy and Efficiency in ASR with Diffusion-Based Non-Autoregressive Decoding, Hao Yen et al., Proceedings of Interspeech 2026 (arXiv:2602.18952), 2026, Table 3 (English benchmarks), MDM-ASR (ours) row.
Open the original (DOI 10.21437/Interspeech.2026-488)- What it measures
- The paper's RTFx is the total duration of audio divided by the total decoding time. The runtime results are on the LibriSpeech test-clean and test-other sets, averaged over all utterances, in full precision without model compilation or runtime optimization. The system is initialized from NVIDIA's canary-1b-flash and uses a masked diffusion decoder with parallel token prediction.
- What it does not show
- Batch size 1 on one A100 with no runtime optimization, so it is not a batched or production figure. It measures decoding speed on LibriSpeech only. The paper's table marks some baseline rows as taken from other papers; those rows are not used here.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as MDM-ASR, Yen et al. 2026: 47 times faster than real time.
- System
- MDM-ASR (Yen et al. 2026)
- Kind
- Research system
- Hardware
- NVIDIA A100 GPU
- Test set
- LibriSpeech test-clean and test-other
- Year
- 2026