基于BLSTM-CTC的语音特征的音素识别研究  

Research on Phoneme Recognition Based on Speech Features

作　　者：吴丹丹 夏秀渝[1] Wu Dandan;Xia Xiuyu(College of Electronic and Information Engineering,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学电子信息学院,成都610065

出　　处：《现代计算机》2022年第10期32-38,共7页Modern Computer

摘　　要：音音素是自然语言中的最小建模单元,音素识别模型的优劣直接影响关键词检索、连续语音识别的性能。本文首先针对幅度特征MSRCC和相位特征PSRCC进行了一系列对比实验研究,发现融合幅度特征和相位特征可以取得更好的识别效果;接着比较分析了几种深度神经网络的优缺点,并将它们用于音素识别,仿真实验表明基于BLSTM-CTC的声学模型相比于其他模型具有更好的识别性能。Phoneme is the smallest modeling unit in natural language,and the quality of phoneme recognition model directly affects the performance of keyword retrieval and continuous speech recognition.This paper firstly conducts a series of comparative experimental studies on the amplitude feature MSRCC and the phase feature PSRCC,and finds that the fusion of the amplitude fea⁃ture and the phase feature can achieve better recognition results;For phoneme recognition,simulation experiments show that the acoustic model based on BLSTM-CTC has better recognition performance than other models.

关 键 词：音素识别 深度神经网络 语音特征 

分 类 号：TN912.34[电子电信—通信与信息系统]