Recognition of unknown words in direct acoustic word speech recognition using acoustic word embedding

摘要

A method (and structure and computer) for learning unknown (OOV) words in an automatic speech recognition (ASR) system is an acoustic word embedding recurrent neural network (AWE RNN) that receives a new OOV word sequence for the ASR system. ), The RNN provides an acoustic word embedding (AWE) vector as its output. The AWE vector output from the AWE RNN is provided as input to the acoustic word embedding-acoustic word neural network (AWE → A2W NN) trained to provide OOV word weights from the AWE vector. The OOV word weights are inserted into the list of acoustic word (A2W) word embeddings used by the ASR system to output the words recognized from the input of audiovisual features, and the OOV word weights are the A2W word embedding list. It is inserted into the A2W word embedding list compared to the existing weights in.