I was able to do it by having VoiceInterfaceSteam use FVoiceEngineImpl instead of it’s own voice engine; the only other main change was using unique net ids converted to strings instead of directly for mapping the talkers (the ids are an opaque type).