Multilingual connected digits and natural numbers recognition in the telephone speech dialog systems

This paper presents the system for multilingual connected digits recognition and natural numbers recognition over the telephone. System bases on language dependent HMM recognisers operating in parallel and is currently implemented for the Slovene and German language. It has been developed with the HTK V2.1.1 with improved frontend processing module. The language dependent recognisers were trained using the Slovene and German SpeechDat(II) databases. The system was tested in both a monolingual and in a multilingual environment in the off-line and on-line mode of operation. For off-line testing the speech files from the Slovene and German SpeechDat(II) databases were used, while for on-line testing the system was integrated into a Reverse Directory System O-TEL (a real world environment). During the testing of the system, various HMM architectures were tested (whole-word models, phone models, triphone models) in order to find the recogniser configuration that is the most efficient in a multilingual environment. Experiments have shown the high recognition accuracy in a single-language environment as well as in a multilingual environment. The multilingual tests have also pointed out the high language identification rate of the multilingual system. A number of advantages of multilingual speech recognition in speech dialog systems were found while experimenting with the O-TEL system.