Nvidia releases an open speech model that transcribes 40 languages