Telephone speech recognition using simulated data from clean database

Abstract

Speech recognition over lines forms an integral part of various applications of large vocabulary continuous speech recognition (LVCSR). This paper describes an implementation system completely in software form to produce simulated telephone data starting from clean databases. Filters adopted in this system are well-designed to simulate the frequency properties of analogue transmission equipments in telephone connection. A speech recognizer was trained from speech data extracted from clean corpus piped through a hardware simulator. The recognition performances are evaluated on a real telephone speech set and several test sets simulated from the clean database for test use. The experiments verified the effectiveness and feasibility of software simulation from a recognition testing point of view, and the results showed using simulated data derived from clean corpus could achieve the same recognition performance as real telephone speech.