Voice phoneme recognition method based on consistency router network

• Main Authors: XU NING, NI YANAN, LIU XIAOFENG, YAO XIAO
• Format: Patent
• Language: Chinese; English
• Published: 10.05.2019
• Subjects: ACOUSTICS; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

The invention discloses a voice phoneme recognition method based on a consistency router network. The voice phoneme recognition method comprises the following steps: configuring a consistency router feed-forward network encoder at a training stage, and performing matrix operations and two-dimensional convolution on initial vector voice phoneme data so as to obtain a low-order state vector; configuring a consistency router structure of the consistency router network, weighting the low-order state vector by virtue of a consistency router coefficient, and transferring the information therein to high-order state vectors; representing and determining probability of the label by virtue of a ratio of a softmax activation value of the length of each high-order state vector to softmax activation values of all the high-order state vectors; configuring a reconstruction decoder of the consistency router network, and performing decoding reconstruction on a high-order state vector corresponding to areal label; performing o