摘要
传统的语音智能识别系统对低资源语言的识别能力较差,识别结果准确性低。为解决这一问题,基于长短时记忆元设计了一种新的语音智能识别系统,针对系统硬件和软件进行优化。选用SPCE061A单片机作为系统硬件的核心处理器,并在硬件内部构建了LSTM-HMM声学模型,通过I/O接口连接各单元模块,无线传输模块内部核心芯片为nRF2401A芯片。软件程序由过构建声学模型、提取语言特征、监督训练语言文本、判断语速、语音识别优化等步骤完成,利用神经元网络设计了语音特征提取程序。为检测设计的识别系统工作效果,与传统识别系统进行实验对比,结果表明,相较于传统系统,基于长短时记忆元设计的语音智能识别系统对低资源词汇的识别错误率下降了29.3%,查询的关键词权重代价提升了71.2%。
Traditional speech intelligent recognition system has poor recognition ability for low-resource languages and low accuracy of recognition results. To solve this problem,a new intelligent speech recognition system based on long-term and short-term memory elements is designed,and the hardware and software of the system are optimized. SPCE061 A MCU is selected as the core processor of the system hardware,and LSTM-HMM acoustic model is built in the hardware. Each unit module is connected through I/O interface. The core chip of the wireless transmission module is nRF2401 A chip. The software program is completed by constructing acoustic model,extracting language features,supervising and training language text,judging speech speed,speech recognition optimization and so on. The speech feature extraction program is designed by using neural network structure. In order to test the performance of the recognition system,the experiment results show that compared with the traditional system,the recognition error rate of the speech intelligent recognition system based on long-term and short-term memory elements is reduced by 29.3%,and the cost of keyword weight is increased by 71.2%.
作者
杨志杰
张梅
李冠龙
黄昌达
YANG Zhi⁃jie;ZHANG Mei;LI Guan⁃long;HUANG Chang⁃da(State Grid Xinjiang Electeic Power Research Institute CO.,LTD,Urumqi 830000,China)
出处
《电子设计工程》
2020年第1期55-58,64,共5页
Electronic Design Engineering
基金
国家电网公司科技项目(PD71-17-008)
关键词
长短时记忆元
语音识别
智能识别
语音识别系统
声学模型
long-term and short-term memory element
speech recognition
intelligent recognition
speech recognition system
acoustic model
