基于CTC-GRU模型的长沙方言识别

Changsha Dialect Recognition Based on CTC -GRU Model

为了识别大词汇量下连续长沙话方言语音,提出了基于CTC算法的门控线性单元神经网络模型.先通过梅尔倒谱系数提取语音的特征参数,再把提取的特征参数输入门控线性单元神经网络,用CTC算法进行训练优化,得到输入序列整个的预测标签.最后在自建的长沙话方言语料库上,以词错率作为评价指标,对CTC模型、GRU模型和CTC-GRU模型进行对比,结果表明CTC-GRU模型相对于其他2个模型收敛速度更快,结果更精准.

In order to recognize continuous speech in Changsha dialect with a large vocabulary,a gated linear element neural network model based on Connectionist Temporal Classification(CTC)algorithm is proposed.Firstly,the characteristic parameters of speech are extracted by Mel-scale Frequency Cepstral Coefficients(MFCC),and then the extracted characteristic parameters are input into gated linear unit neural network.CTC algorithm is used for training and optimization,and the whole prediction label of input sequence is obtained.Finally,the results of the CTC model,the GRU model and the CTC-GRU model are compared on the self-built corpus of Changsha dialect,and the Word Error Rate(WER)is taken as the evaluation index.The results show that the CTC-GRU model can achieve faster convergence and greater accuracy compared with the other two models.