Offline Speech Translation on Jetson Nano Using Quality-Aware Correction
October 7, 2026
A fully offline speech-to-speech pipeline for Jetson Nano (4 GB) uses Whisper-tiny and Opus-MT with a multilingual BERT quality gate. Implementing Minimum Bayes-Risk decoding at a 0.90 threshold improved English-Spanish translation by +0.67 BLEU and +0.51 ChrF.
HOW THIS AFFECTS YOU
●
builderYou can deploy high-quality, self-correcting translation pipelines on low-memory edge devices.
●
researcherThe findings suggest that Quality Estimation works well as a gate but is ineffective for reranking.