This paper presents two real-time systems for recognizing sentence-level continuous American Sign Language (ASL) using a single camera to track the user's unadorned hands. The first system uses a desk-mounted camera and achieves 92% word accuracy, while the second system, which uses a camera mounted on a cap worn by the user, achieves 98% accuracy (97% with an unrestricted grammar). Both systems use a 40-word lexicon. ASL is a structured language with specific gestures and rules for context and grammar. The systems described in this paper use hidden Markov models (HMMs) for recognition. HMMs are well-suited for recognizing complex, structured hand gestures as found in sign languages. The systems do not require explicit segmentation on the word level for training or recognition. Language and context models can be applied at various levels, and much related development has been done by the speech recognition community. The first system, a desk-based recognizer, uses a single color camera to track unadorned hands in real time and interpret ASL using HMMs. The second system, a wearable-based recognizer, uses a camera mounted on a cap worn by the user. Both systems use a single camera to track unadorned hands in real time and interpret ASL using HMMs. The wearable system is more accurate than the desk system due to less occlusion and better data recording methods. The systems use a 40-word lexicon and recognize sentences of the form "personal pronoun, verb, noun, adjective, (the same) personal pronoun." The systems use HMMs to recognize the signed words. The first system achieves 92% word accuracy, while the second system achieves 98% accuracy. The systems use HMMs to recognize the signed words. The wearable system is more accurate than the desk system due to less occlusion and better data recording methods. The systems use a 40-word lexicon and recognize sentences of the form "personal pronoun, verb, noun, adjective, (the same) personal pronoun." The systems use HMMs to recognize the signed words.This paper presents two real-time systems for recognizing sentence-level continuous American Sign Language (ASL) using a single camera to track the user's unadorned hands. The first system uses a desk-mounted camera and achieves 92% word accuracy, while the second system, which uses a camera mounted on a cap worn by the user, achieves 98% accuracy (97% with an unrestricted grammar). Both systems use a 40-word lexicon. ASL is a structured language with specific gestures and rules for context and grammar. The systems described in this paper use hidden Markov models (HMMs) for recognition. HMMs are well-suited for recognizing complex, structured hand gestures as found in sign languages. The systems do not require explicit segmentation on the word level for training or recognition. Language and context models can be applied at various levels, and much related development has been done by the speech recognition community. The first system, a desk-based recognizer, uses a single color camera to track unadorned hands in real time and interpret ASL using HMMs. The second system, a wearable-based recognizer, uses a camera mounted on a cap worn by the user. Both systems use a single camera to track unadorned hands in real time and interpret ASL using HMMs. The wearable system is more accurate than the desk system due to less occlusion and better data recording methods. The systems use a 40-word lexicon and recognize sentences of the form "personal pronoun, verb, noun, adjective, (the same) personal pronoun." The systems use HMMs to recognize the signed words. The first system achieves 92% word accuracy, while the second system achieves 98% accuracy. The systems use HMMs to recognize the signed words. The wearable system is more accurate than the desk system due to less occlusion and better data recording methods. The systems use a 40-word lexicon and recognize sentences of the form "personal pronoun, verb, noun, adjective, (the same) personal pronoun." The systems use HMMs to recognize the signed words.