In　the　field　of　instant　voice　communication　(IVC)　steganalysis,　the　traditional　detecting　methods　are　mainly　based　on　supervised　learning　scheme　that　results　in　a　large　amount　of　complex　manual　pre-processing　training　data　set.　The　accuracy　of　supervised　learning　scheme　can　be　easily　destroyed　by　the　difference　between　the　distribution　of　training　and　testing　data　set　in　the　actual　voice　application.　The　disadvantages　of　this　method　are　obvious　in　the　big　data　environment.　In　this　regard,　this　paper　initially　introduced　a　novel　semi-supervised　hybrid　learning　detection　model　for　the　IVC　network.　This　provides　the　progress　of　manually　annotating　training　data　set　that　has　been　removed　to　solve　the　problem　of　complex　operations　and　poor　applicability　in　classifier.　Therefore,　this　model　has　a　simpler　structure　and　more　extensive　detection　scopes　with　the　huge　amount　of　data.　Then,　we　designed　a　multi-criteria　fusion　module　that　can　automatically　generate　the　pseudo-label　set　from　testing　data　set　to　train　the　classifier　model.　Thus,　our　scheme　will　not　be　affected　by　the　distribution　shift.　In　this　module,　we　defined　the　confidence　level　and　representative　level　to　judge　the　feature　vector　for　pseudo-labeled.　Through　the　experimental　analysis,　the　low　bit-rate　speech　coding　steganalysis　(G.723.1/G.729/iLBC　speech　codecs)　is　analyzed　on　quantization　index　modulation　that　are　common　codecs　in　the　IVC　network.　The　results　show　that　our　method　has　higher　accuracy　than　un-supervised　method.　The　proposed　approach　is　less　affected　and　more　accurate　than　the　previous　supervised　methods　through　the　distribution　of　different　training　and　testing　data　sets.　The　experiments　also　proved　that　our　method　can　be　deployed　in　the　different　kinds　of　the　IVC　codec　by　considering　huge　amount　of　data　set.