基于改进型轻门控循环单元的语音识别  

Improved light gated recurrent unit for speech recognition

作　　者：蒋畅江[1] 温登峰 唐贤伦[1] 巴莉芳 JIANG Chang-jiang;WEN Deng-feng;TANG Xian-lun;BA Li-fang(School of Automation,Chongqing University of Posts and Telecommunication,Chongqing 400065,China)

机构地区：[1]重庆邮电大学自动化学院

出　　处：《计算机工程与设计》2019年第11期3265-3268,3356,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61673069)

摘　　要：针对轻门控循环单元(light gated recurrent unit,Li-GRU)在自动语音识别中存在参数规模大、收敛速度慢和梯度消失的问题,提出一种改进的模型(light self gated recurrent unit,Li-SGRU)。使用Swish激活替换ReLU激活,使得模型在训练时能够避免梯度消失,加快网络收敛;提出3种变体Li-SGRU1、Li-SGRU2和Li-SGRU3减少改进模型的参数规模;在端到端的语音识别系统上评价了该模型。实验结果表明,改进的Li-SGRU较标准的模型音素错误率更低,收敛速度更快;其中,Li-SGRU1识别率在3种变体中最高,综合性能更好。Aiming at the problem of large parameter size,low convergence speed and disappearing gradient in automatic speech recognition of light gated recurrent unit(Li-GRU),an improved model(light self gated recurrent unit,Li-SGRU)was proposed.Swish activation was used to replace ReLU activation,so that the model could avoid gradient disappearance and accelerate network convergence during training.Three variants Li-SGRU1,Li-SGRU2 and Li-SGRU3 were proposed to reduce the parameter size of the improved model.The proposed model was evaluated on the end-to-end speech recognition system.Experimental results show that the improved Li-SGRU has lower phoneme error rate and faster convergence than the standard model.Moreover,the recognition rate of Li-SGRU1 is the highest among the three variants,and its comprehensive performance is better.

关 键 词：轻门控循环单元 自动语音识别 Swish激活 ReLU激活 端到端 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]