Following segmentation and phonetic classification in automatic recognition of continuous speech (ARCS), it is necessary to provide methods for linguistic decoding. In this work a graph search procedure, based on the Fano algorithm, is used to convert machine-contaminated phonetic descriptions of speaker performance into standard orthography. The information utilized by the decoder consists of a syntax, a lexicon containing transcription variation for each word, and performance-based statistics from acoustic analysis. The latter contain information related to automatic segmentation and classification accuracy and certainty (anchor-point) data. A distinction is made between speaker- and machine-dependent corruption of phonetic input strings. Preliminary results are presented and discussed, together with some considerations for evaluation. © 1973, IEEE. All rights reserved.