Peer-reviewed research
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 · DOI 10.21437/Interspeech.2026-488
Read the originalRead on October 19, 2026
- What it is
- A research paper from Interspeech 2026 proposing a non-autoregressive speech recognizer built on masked diffusion, initialized from NVIDIA's canary-1b-flash. Its results table reports word error rate and inverse real-time factor (RTFx) for its own system and for baselines, and states that the RTFx runs were done on a single NVIDIA A100 GPU with a batch size of 1.
- How much weight it carries
- Peer-reviewed conference paper with an open full text on arXiv. The RTFx figure is for the authors' own system on LibriSpeech test-clean and test-other. Some baseline rows come from other papers and are not measured on the same hardware, so only the authors' own row is used here. The paper measures decoding speed only for English LibriSpeech audio.
Facts taken from this source
- 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.
“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” (Table 3 (English benchmarks), MDM-ASR (ours) row)