An experiment in word hypothesization performed in the context of a continuous speech recognition system

The experiment consists in testing an hypothesizer that assembles, in different degrees of corruption of the input phonetic sequence, data on the processing time, on the complexity of those underlattices which develop "around" a single lexical hypothesis as well on performances, in numerical terms, of the separate hypotheses of the words present in the symbolic input sequence. An ltalian sentence, in graphernic form, is phonetically transcribed and it undergoes a phase of corruption in which every space is suppressed and errors of insertion, deletion and Substitution are introduced in order to reach an deterrnined error rate. The corrupting program makes use of statistical knowledge on the performance of acousticalphonetic decoder and on experiments about human perception. The sequence so corrupted is presented tothe lexical hypothesizer.