Fact · Speech recognition
5.99times faster than real time
The best fine-tuned Whisper-small model, NOMIMOSE, processed audio at 5.99 times real time on an NVIDIA Jetson, while also having the lowest pooled word error rate of 22.72 percent.
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“Its audio throughput reached 5.99×real time, with an inference throughput of 3.50”
Fairness Evaluation of Edge-AI Implementation for Cleft Lip and Palate Speech ASR, Susmita Bhattacharjee et al., arXiv, 2026, Section 3.1, Deployment Results.
Open the original- What it measures
- Fine-tuned on normal and mild, moderate and severe cleft speech, then averaged over five folds and run on the Jetson GPU in FP16.
- What it does not show
- Same single edge deployment as the baseline row. The paper gives this as audio throughput, which is the inverse of its real-time factor of 0.167.
In the data
A row of Speech Recognition Speed: Real-Time Factor Data, as Bhattacharjee et al. 2026, NOMIMOSE: 6 times faster than real time.
- System
- Whisper-small, NOMIMOSE fine-tuned
- Kind
- Open model
- Hardware
- NVIDIA Jetson (ARM CPU + CUDA GPU)
- Test set
- NMCPC cleft lip and palate speech, 264 utterances
- Year
- 2026