计算机应用与软件2023，Vol.40Issue(12)：184-188,262,6.DOI:10.3969/j.issn.1000-386x.2023.12.027
基于QRNN-CTC的中文语音识别声学模型
CHINESE SPEECH RECOGNITION ACOUSTIC MODEL BASED ON QRNN-CTC
摘要
Abstract
Aimed at the problem of insufficient processing time sequence ability of convolutional neural network(CNN)in speech recognition and high model complexity and difficulty of training in recurrent neural network(RNN)in speech recognition,a new kind of quasi-recurrent neural network and connectionist temporal classification(QRNN-CTC)acoustic model is proposed.It not only reduced the numbers of parameters but also ensured a certain cycle capability between time series.CTC was used to realize automatic alignment of input sequence and label,and dropout was introduced to prevent overfitting during training.The experimental results on the Thchs-30 dataset show that QRNN-CTC has a relative error rate of 9.8%lower than that of CNN-CTC,and the final word error rate is23.8%,and the training time is half of LSTM-CTC.关键词
深度学习/语音识别/声学模型/准循环神经网络/连接时序分类Key words
Deep learning/Speech recognition/Acoustic model/QRNN/CTC分类
信息技术与安全科学引用本文复制引用
王先欢,孙自强..基于QRNN-CTC的中文语音识别声学模型[J].计算机应用与软件,2023,40(12):184-188,262,6.基金项目
中央高校基本科研业务费专项资金资助项目(222201917006). （222201917006）
