Fact · Speech recognition
1120.88times faster than real time
In the Galvez et al. 2024 paper, Parakeet RNN-T 1.1B with label looping and CUDA graphs ran at an RTFx of 1120.88 on an A100-80GiB PCIe GPU at batch size 32, meaning about 1,121 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
“Label Looping + CUDA Graphs 1120.88 17.52 ... For all results, we use an A100-80GiB PCIe card, and a batch size of 32”
Speed of Light Exact Greedy Decoding for RNN-T Speech Recognition Models on GPU, Daniel Galvez et al., Proceedings of Interspeech 2024 (arXiv:2406.03791), 2024, Section 4 (Results) and Table 1, Parakeet RNN-T 1.1B, Label Looping + CUDA Graphs row.
Open the original (DOI 10.48550/arXiv.2406.03791)- What it measures
- The paper defines RTFx as the number of hours an ASR model can transcribe in one hour of wall-clock time. It is an offline inference task on LibriSpeech test-other (5.1 hours), with bfloat16 inference. In Table 1 this is the fastest RNN-T 1.1B decoder that keeps the baseline WER of 2.7 percent.
- What it does not show
- Offline and batched at batch size 32, with bfloat16. This is not a batch-size-1 or single-file speed, so it is not directly comparable with the batch-1 A100 figure from the MDM-ASR paper. The number is the decoder's speed on NVIDIA's open Parakeet checkpoint.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as Galvez et al. 2024, looping + graphs: 1,121 times faster than real time.
- System
- Parakeet RNN-T 1.1B, label looping and CUDA graphs
- Kind
- Research system
- Hardware
- NVIDIA A100-80GiB PCIe GPU
- Test set
- LibriSpeech test-other, offline, batch size 32
- Year
- 2024