Vocabulary-independent recognition of american Spanish phrases and digit strings

We describe the development of an R&D recognizer for several Spanish applications, starting from an existing recognition system for American English and modest language-speci c resources. The experiments emphasize achieving phonetic accuracy on telephone speech without vocabulary speci c training. We use our basic recognition engine, and simple grammar-building tools for predicting word sequences. Only the read sentences from two telephone speech corpora (Voice Across Hispanic America (VAHA) and a smaller TI corpus) are used for training. Word error rates (WER) of 1.9% on telephone service command phrases, 5.5% on telephone numbers, and 12% on continuously spoken sentences are achieved with the newly ported system.