A connected speech recognition system using a diphone-based language model

Almost all CSR systems presently in practical use are based on whole-word template matching. Although their performances are quite high, a few problems arise due to the use of whole words as ba sic units. They are related to storage require ments, coarticulation effects at the junction be tween words, discrimination of similar words and training time, especially for large vocabularies. In this paper we report a very positive experi ence which is being made with a system based on very short sub-word units, called "diphones". With the approach described in this paper, the above mentioned problems are much alleviated, with no penalty on performance. The first part is devoted to the presentation and discussion of the peculiar characteristics of the diphones, of the language model based on them and of the overall recognition system. Then a set of procedures used for training the system on a new application and for extracting the diphone templates for any new speaker are briefly de scribed. Finally we report and discuss the results of different tests performed on various recogni tion tasks.