融合CTC/Attention联合解码的拉萨方言抗噪语音识别研究

打开文本图片集

中图分类号：TP319 文献标识码：A

Abstract：LhasaDialectisanimportantbranchofTibetanlanguage，andconductingresearchonspeechrecogntionofLhasa Dialectholdssignificanttheoreticalandpracticalvalue.AddresingtheissuesofcarceresearchonnoisyLhasaDialectspeechand highWordErorRate（WER），thisstudyintegratesspeehenhancementalgorithmsandspeechrecognitionmodelsintoanoiserobustspechrecognitionsystemforLhasaDialectMeanwhile，itimprovestheConformerspeechecognitionmodelbyintroducing theCTC/tentionjointdecodingmechanism，formingtheConformer-CTC/Atentionspeehrecognitionmodeltofurtherreduce WER.ExperimentalresultsshowthattheconstructedConformer-CTC/AtentionspechrecognitionmodelachievesaWERof 25.46% onthepublicTibetanlanguagespeechdataset，whichisgenerallylowerthanthatof mainstreammodels.Furthermore，the WERof the constructed noise-robust speech recognition model forLhasa Dialect on noisyspeech is reduced by 8.9% compared with that of non-noise-robust speech recognition models.

Keywords：Lhasa Dialect;Speech Recognition; Conformer;Speech Enhancement

0引言

方言不仅是藏语的重要分支，也是广播电视领域的通用语言，在藏语文化传播与日常语音交互中占据核心地位。（剩余8162字）