Abstract: We describe a simple speech translation architecture intended for medical and other safety-critical applications, which is intended to represent a compromise between fixed-phrase translation on one hand and complex transfer-based translation on the other. Recognition is guided by an annotated CFG-based language model compiled from a unification grammar; transfer and generation use a minimal list-oriented semantic representation language. We present an evaluation of an initial prototype, which translates yes/no questions about hypoglycaemia from spoken French into spoken English using a vocabulary of about 200 words.
0 Replies
Loading