LSTM网络在语音转文字应用中的优化方法  

Optimization Method of LSTM Network in Speech-to-Text Application

作　　者：张乾[1] ZHANG Qian(Jianghai Polytechnic College,Yangzhou 225101,China)

机构地区：[1]江海职业技术学院,江苏扬州225101

出　　处：《电声技术》2024年第9期85-87,共3页Audio Engineering

摘　　要：为研究基于长短期记忆(Long Short-TermMemory,LSTM)网络的语音转文字系统的优化方法,首先说明LSTM在语音转文字任务中的基本原理和架构,其次分析自适应矩估计(Adaptive Moment Estimation,Adam)优化算法的核心机制及其在LSTM网络中的应用,最后在Mozilla DeepSpeech框架中嵌入基于Adam优化的LSTM模型,并使用THCHS-30数据集进行实验。实验结果表明,基于Adam优化的LSTM模型在词错率和F1分数上均表现出显著的优越性。In order to study the optimization method of speech-to-text system based on Long Short-Term Memory(LSTM)network,the basic principle and architecture of LSTM in speech-to-text task are first explained,and then the core mechanism of Adaptive Moment Estimation(Adam)optimization algorithm and its application in LSTM network are analyzed.Finally,the LSTM model based on Adam optimization is embedded in the Mozilla DeepSpeech framework,and the experiment is carried out using the THCHS-30 dataset.The experimental results show that the LSTM model based on Adam optimization has obvious advantages in terms of word error rate and Ff score.

关 键 词：长短期记忆(LSTM) 自适应矩估计(Adam) 语音识别 训练优化 

分 类 号：TN912.3[电子电信—通信与信息系统]