Sound　source　detection　and　localization　have　a　lot　of　practical　uses　in　many　industrial　settings.　Most　of　sound　source　direction　detection　algorithms　in　literature　are　designed　to　identify　the　angle　of　sound　source　in　a　2D　space.　In　this　work,　we　propose　to　use　convolutional　neural　networks　to　detect　the　sound　source　direction　in　a　3D　space.　This　algorithm　is　based　on　the　generalized　cross　correlation　method　with　phase　transform　(GCC-PHAT)　[1]　to　derive　time　delay　of　arrival　(TDOA).　By　using　a　convolutional　neural　network　model,　this　algorithm　can　be　applied　and　deployed.　In　addition,　by　modifying　GCC-PHAT　formula,　this　approach　also　works　of　multiple　sound　sources　detection.　Simulation　experimental　results　on　single　sound　source　and　multiple　sound　sources　detection　show　the　proposed　system　could　work　in　most　situations.　©　2018　IEEE.