This paper describes the implementation and evaluation of a voice-activated dialling (VAD) system. The system allows an open and customised vocabulary set. The pronunciation of the enrolled entry is produced by the system from spoken examples provided by the users. The recognition rate of the baseline system is 96.43% with vocabulary size of 52. System performance of both single utterance (SU) training and multiple utterance (MU) training approaches are evaluated. Various methods to implement the MU training in VAD is proposed. The system error rate can be reduced by up to 72% when MU training is used. If the system is moved from single user to multiple users environment where users other than the enrolled speaker are also allowed to use the same system without re-enrolment, recognition rate of up to 94.64% can be achieved with the application of MU training.
展开▼
