Search International and National Patent Collections
|1. (WO2017177484) VOICE RECOGNITION-BASED DECODING METHOD AND DEVICE|
|Applicants:||AISPEECH CO., LTD.
SHANGHAI JIAO TONG UNIVERSITY
|Title:||VOICE RECOGNITION-BASED DECODING METHOD AND DEVICE|
A voice recognition-based decoding method and device. The method comprises: receiving voice information to extract an acoustic feature (S101); computing information about the acoustic feature according to a connectionist temporal classification model (S102); if a frame in the information about the acoustic feature is a non-empty model frame, searching linguistic information using a weighted finite state transducer adapted to acoustic modelling information and storing historical information, or otherwise, discarding the frame (S103). By constructing a connectionist temporal classification model, acoustic modeling becomes more accurate. By using an improved weighted finite state transducer, model representation becomes more efficient, thereby reducing 50% of the computation and the memory resource consumption. By using a phoneme synchronization method during decoding, the complexity and the number of times of computation during model searching are effectively reduced.