Segmenting discrete data representing continuous speech input

摘要

A probabilistic method for segmenting continuous speech into lexical units is described. The algorithm assumes initial conversion of the continuous speech signal to a discrete representation over some suitable alphabet. The problem of determining such alphabets is not considered. Experiments used keyed input in English, French, German, and Russian. We hypothesize that the low error rates obtained in the experiments can also be achieved with data representing actual speech. The paper discusses an area of linguistic science, and outlines a method for investigating it.