To enable translations for low-resource languages & to prep for future real-time speech to speech translations, we’re expanding our automatic data set creation techniques, working to overcome modeling challenges, and finding new ways to evaluate MT results.