基于DRSN-Conformer的电力调度语音识别

打开文本图片集

中图分类号：TN912.3-34 文献标识码：A 文章编号：1004-373X（2026）06-0112-08

Power dispatching speech recognition based on DRSN-Conformer

HAN Yaxu1，GAO Lu1， ZHANGFei²，QIN Ling1 ，WANG Yongping1， ZHANG Xiaolin1 o 2.SchloftomatinandEletricalEineeingrogoliaUvesityfieneandholootou4）

Abstract：Inallusion totheproblemsof Transformer network-basedspeech recognitionmodels，such asinsuficientability toextractaudiofeaturesinpowerdispatchingnoisescenarios，lowaccuracyinrecognizingprofesional terms，andpoor robustness in powerdispatch noise scenarios，a powerdispatching scenariospeech recognition methodcombining deepresidual shrinkagenetwork（DRSN）andConformer network isproposed.InDRSN，twotypesofcontractionmodules，namelychanel thresholdsharingmoduleandindependentshrinkagemodule，aredesignedtolearnthenoisethresholdsofaudiosignals，thereby reducinginterferencecausedbynoise.TheaudiosignaloutputbytheDRSNblock isencoded intoaudiofeaturesbymeansof Conformerblock，andthecross-attentionisusedtofusethetextfeaturesencodedbythetextencodertoobtaintherelevance representationofaudioandtext.Themodelistrainedanddecodedbycombining withtheCTClossandAtentionloss.The experimentalresultsonthepublicdatasetsAshell-1，Thchs3O，andaninteralpowerdispatchdatasetofapowergridcompany inInerMongoliademonstratethat，incomparisonwiththeTransformermodel，thereisan8.5%decreaseintheultimate charactererrorrate，and5.2%decreaseinthecharactererorateofspecialtermrecognition，validatingtheefectivenessand advantages of the proposed method in solving the power dispatch speech recognition tasks.

Keywords： speech recognition; DRSN; Conformer; power dispatching; attention mechanism; BiLSTM

0 引言

随着各行各业电力需求的日益增加，电力系统调度工作愈发繁重。（剩余13331字）