基于深度神经网络的成都话识别研究

Chengdu Dialect Recognition Based on Convolutional Neural Network

随着智能语音技术的普及和各地文化的互相影响,使用方言进行人机语音交互已经成为了新的研究方向。为了在本土影视剧里能更有效标注对应的普通话字幕,针对成都话的发音特点,提出一种基于卷积神经网络的成都话识别方法,通过自建成都话语料库,构建由CNN、CTC组成的识别模型,实现了从成都话到普通话的映射识别研究。实验结果表明,使用文中方法错词率最终能降到3.23%。

Following the popularization of intelligent speech technology and the mutual influence of local cultures,the use of dialects for human-machine speech interaction has become a new research direction.In accordance with the pronunciation characteristics of Chengdu dialect,a recognition method of Chengdu dialect based on convolutional neural network is proposed to mark the corresponding Mandarin subtitles more effectively in local television dramas.A recognition model composed of CNN and CTC is constructed by the establishment of Chengdu dialect corpus to study the mapping and recognition from Chengdu dialect to Mandarin.The experimental results show that the error rate is reduced to 3.23%by the method suggested in the paper.