Preprint
Fairness Evaluation of Edge-AI Implementation for Cleft Lip and Palate Speech ASR
Susmita Bhattacharjee et al., arXiv, 2026
Read the originalRead on October 19, 2026
- What it is
- A preprint that fine-tunes Whisper-small on normal and cleft lip and palate speech, then deploys the models on an NVIDIA Jetson board. It reports word error rates and real-time factors for the pretrained and fine-tuned models on 264 evaluation utterances.
- How much weight it carries
- ArXiv preprint, not yet peer reviewed. Speed figures come from one Jetson deployment on 264 utterances, with greedy decoding and FP16 inference. The hardware is named only as NVIDIA Jetson, without a board model.
Facts taken from this source
- 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.
“Its audio throughput reached 5.99×real time, with an inference throughput of 3.50” (Section 3.1, Deployment Results)
- On an NVIDIA Jetson, the pretrained Whisper-small model had a real-time factor of 0.212, which is 4.72 times faster than real time, on 264 cleft lip and palate utterances.
“The pretrained Whisper-small baseline achieved an overall RTF of 0.212, corresponding to4.72×real-time processing” (Section 3.1, Deployment Results)