Language Recognition Test and Evaluation.

Abstract
A language classification algorithm was developed that uses multiphoneme reference sounds and a summed log-likelihood decision rule with rejection incorporated. The multilanguage data base was expanded from five to seven languages. This data base contains segments (10-15 minutes in duration) of text read by adult males. A collection of 80 reference sounds was generated from 66 design speakers of the seven available languages. These references were used in a 65-speaker, 7-language evaluation, where 100% correct classification was obtained using a rejection threshold that yielded 68% rejection. A five-language test indicated that 56% rejection provided zero classification errors. A data processor simulation was implemented that incorporated automatic detection of onset and cessation of speech, along with concurrent aural playout of either the incoming data or operator-selected, prestored data. (Author)

This publication has 0 references indexed in Scilit: