Early　intervention　for　depression　could　provide　a　means　to　reducing　the　disease　burden,　but　there　is　a　lack　of　objective　diagnostic　methods.　This　study　investigated　automatic　depression　classification　on　a　speech　dataset　of　85　healthy　controls　(51　females　and　34　males)　and　85　depressed　patients　(53　females　and　32　males).　Considering　that　there　are　obvious　differences　in　the　performance　of　different　types　of　speech　features,　we　propose　a　radius-incorporated　localized　multiple　kernel　learning　(trLMKL)　algorithm　for　detecting　depression　in　speech　to　make　the　best　use　of　speech　features.　To　improve　the　classification　accuracy,　we　combine　the　information　of　both　the　margin　and　the　radius　of　the　MEB　to　learn　the　gating　model　parameters　in　our　algorithm.　Furthermore,　we　do　not　directly　incorporate　the　radius　of　the　MEB,　but　incorporate　the　trace　of　the　total　scattering　matrix　of　training　data.　This　method　can　avoid　the　time　cost　of　calculating　the　radius　at　each　iteration　and　decrease　the　computational　complexity.　Comprehensive　experiments　were　carried　out　on　our　depressed　speech　dataset　and　10　UCI　datasets.　Our　algorithm　achieved　better　classification　performance　overall　than　SimpleMKL　and　LMKL,　and　it　was　efficient　at　detecting　depression,　indicating　its　potential　for　use　as　a　diagnostic　method　for　depression.　©　2023,　The　Author(s),　under　exclusive　licence　to　Springer　Science+Business　Media,　LLC,　part　of　Springer　Nature.