Utilization of contextual constraints in automatic speech recognition

The use of contextual constraints in speech recognition has been contemplated by many authors. This paper describes a system under development which will perform correction of errors made in a phonemic recognizer by comparison of the received phoneme sequence with the syntax and dictionary of "the language being spoken. The language syntax is stored in a relatively efficient manner, being essentially in Backus-Naur Form. The error correction procedure can best be described as a sequential decoding on the tree of syntax generated by the tables, as traversed by a syntactical parser. The system is currently being implemented in a form such that it will recognize a "spoken FORTRAN" language. Some initial results of its application to certain error-containing inputs are described.