diff --git a/bob/learn/misc/old/ivector_trainer.cc b/bob/learn/misc/old/ivector_trainer.cc
index e910af9b7b97987165a5ecb14358bb015e3c3856..d98ed5076e02928c92491369f76624c0b2822ea6 100644
--- a/bob/learn/misc/old/ivector_trainer.cc
+++ b/bob/learn/misc/old/ivector_trainer.cc
@@ -80,13 +80,32 @@ static void py_set_AccSnormij(bob::learn::misc::IVectorTrainer& trainer,
   trainer.setAccSnormij(acc.bz<double,2>());
 }
 
+
+
+// include the random API of bob.core
+#include <bob.core/random.h>
+static boost::python::object ITB_getRng(EMTrainerIVectorBase& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+
+#include <boost/make_shared.hpp>
+static void ITB_setRng(EMTrainerIVectorBase& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+
+
 void bind_trainer_ivector()
 {
   class_<EMTrainerIVectorBase, boost::noncopyable>("EMTrainerIVector", "The base python class for all EM-based trainers.", no_init)
     .add_property("convergence_threshold", &EMTrainerIVectorBase::getConvergenceThreshold, &EMTrainerIVectorBase::setConvergenceThreshold, "Convergence threshold")
     .add_property("max_iterations", &EMTrainerIVectorBase::getMaxIterations, &EMTrainerIVectorBase::setMaxIterations, "Max iterations")
     .add_property("compute_likelihood_variable", &EMTrainerIVectorBase::getComputeLikelihood, &EMTrainerIVectorBase::setComputeLikelihood, "Indicates whether the log likelihood should be computed during EM or not")
-    .add_property("rng", &EMTrainerIVectorBase::getRng, &EMTrainerIVectorBase::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
+    .add_property("rng", &ITB_getRng, &ITB_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
     .def("train", &py_train, (arg("machine"), arg("data")), "Trains a machine using data")
     .def("initialize", &py_initialize, (arg("machine"), arg("data")), "This method is called before the EM algorithm")
     .def("finalize", &py_finalize, (arg("machine"), arg("data")), "This method is called at the end of the EM algorithm")
diff --git a/bob/learn/misc/old/jfa_trainer.cc b/bob/learn/misc/old/jfa_trainer.cc
index b530082b79e438b4cc10a0abe00924d43b9b513e..18d0299371a8d3d9635e2978818307aa658d97da 100644
--- a/bob/learn/misc/old/jfa_trainer.cc
+++ b/bob/learn/misc/old/jfa_trainer.cc
@@ -309,12 +309,41 @@ static void jfa_set_accDA2(bob::learn::misc::JFATrainer& trainer,
 }
 
 
+
+// include the random API of bob.core
+#include <bob.core/random.h>
+static boost::python::object isv_getRng(bob::learn::misc::ISVTrainer& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+static boost::python::object jfa_getRng(bob::learn::misc::JFATrainer& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+
+#include <boost/make_shared.hpp>
+static void isv_setRng(bob::learn::misc::ISVTrainer& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+static void jfa_setRng(bob::learn::misc::JFATrainer& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+
+
 void bind_trainer_jfa()
 {
   class_<bob::learn::misc::ISVTrainer, boost::noncopyable >("ISVTrainer", "A trainer for Inter-session Variability Modelling (ISV). \n\nReferences:\n[1] 'Explicit Modelling of Session Variability for Speaker Verification', R. Vogt, S. Sridharan, Computer Speech & Language, 2008, vol. 22, no. 1, pp. 17-38\n[2] 'Session Variability Modelling for Face Authentication', C. McCool, R. Wallace, M. McLaren, L. El Shafey, S. Marcel, IET Biometrics, 2013", init<optional<const size_t, const double> >((arg("self"), arg("max_iterations")=10, arg("relevance_factor")=4.),"Initializes a new ISVTrainer."))
     .def(init<const bob::learn::misc::ISVTrainer&>((arg("self"), arg("other")), "Copy constructs an ISVTrainer"))
     .add_property("max_iterations", &bob::learn::misc::ISVTrainer::getMaxIterations, &bob::learn::misc::ISVTrainer::setMaxIterations, "Max iterations")
-    .add_property("rng", &bob::learn::misc::ISVTrainer::getRng, &bob::learn::misc::ISVTrainer::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
+    .add_property("rng", &isv_getRng, &isv_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
     .add_property("__X__", &isv_get_x, &isv_set_x)
     .add_property("__Z__", &isv_get_z, &isv_set_z)
     .def(self == self)
@@ -333,7 +362,7 @@ void bind_trainer_jfa()
   class_<bob::learn::misc::JFATrainer, boost::noncopyable >("JFATrainer", "A trainer for Joint Factor Analysis (JFA).\n\nReferences:\n[1] 'Explicit Modelling of Session Variability for Speaker Verification', R. Vogt, S. Sridharan, Computer Speech & Language, 2008, vol. 22, no. 1, pp. 17-38\n[2] 'Session Variability Modelling for Face Authentication', C. McCool, R. Wallace, M. McLaren, L. El Shafey, S. Marcel, IET Biometrics, 2013", init<optional<const size_t> >((arg("self"), arg("max_iterations")=10),"Initializes a new JFATrainer."))
     .def(init<const bob::learn::misc::JFATrainer&>((arg("self"), arg("other")), "Copy constructs an JFATrainer"))
     .add_property("max_iterations", &bob::learn::misc::JFATrainer::getMaxIterations, &bob::learn::misc::JFATrainer::setMaxIterations, "Max iterations")
-    .add_property("rng", &bob::learn::misc::JFATrainer::getRng, &bob::learn::misc::JFATrainer::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
+    .add_property("rng", &jfa_getRng, &jfa_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
     .add_property("__X__", &jfa_get_x, &jfa_set_x)
     .add_property("__Y__", &jfa_get_y, &jfa_set_y)
     .add_property("__Z__", &jfa_get_z, &jfa_set_z)
diff --git a/bob/learn/misc/old/kmeans_trainer.cc b/bob/learn/misc/old/kmeans_trainer.cc
index 57129fe4d01f4d210f3271aadf936dcb45031969..3a5838afb121dd15cac78ce5016e89a88344a921 100644
--- a/bob/learn/misc/old/kmeans_trainer.cc
+++ b/bob/learn/misc/old/kmeans_trainer.cc
@@ -56,13 +56,41 @@ static void py_mStep(EMTrainerKMeansBase& trainer,
   trainer.mStep(machine, sample.bz<double,2>());
 }
 
+// include the random API of bob.core
+#include <bob.core/random.h>
+static boost::python::object KMTB_getRng(EMTrainerKMeansBase& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+static boost::python::object KMT_getRng(bob::learn::misc::KMeansTrainer& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+
+#include <boost/make_shared.hpp>
+static void KMTB_setRng(EMTrainerKMeansBase& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+static void KMT_setRng(bob::learn::misc::KMeansTrainer& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+
+
 void bind_trainer_kmeans()
 {
   class_<EMTrainerKMeansBase, boost::noncopyable>("EMTrainerKMeans", "The base python class for all EM-based trainers.", no_init)
     .add_property("convergence_threshold", &EMTrainerKMeansBase::getConvergenceThreshold, &EMTrainerKMeansBase::setConvergenceThreshold, "Convergence threshold")
     .add_property("max_iterations", &EMTrainerKMeansBase::getMaxIterations, &EMTrainerKMeansBase::setMaxIterations, "Max iterations")
     .add_property("compute_likelihood", &EMTrainerKMeansBase::getComputeLikelihood, &EMTrainerKMeansBase::setComputeLikelihood, "Tells whether we compute the average min (square Euclidean) distance or not.")
-    .add_property("rng", &EMTrainerKMeansBase::getRng, &EMTrainerKMeansBase::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
+    .add_property("rng", &KMTB_getRng, &KMTB_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
     .def(self == self)
     .def(self != self)
     .def("train", &py_train, (arg("self"), arg("machine"), arg("data")), "Train a machine using data")
@@ -91,7 +119,7 @@ void bind_trainer_kmeans()
   KMT.def(self == self)
      .def(self != self)
      .add_property("initialization_method", &bob::learn::misc::KMeansTrainer::getInitializationMethod, &bob::learn::misc::KMeansTrainer::setInitializationMethod, "The initialization method to generate the initial means.")
-     .add_property("rng", &bob::learn::misc::KMeansTrainer::getRng, &bob::learn::misc::KMeansTrainer::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of the means.")
+     .add_property("rng", &KMT_getRng, &KMT_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of the means.")
      .add_property("average_min_distance", &bob::learn::misc::KMeansTrainer::getAverageMinDistance, &bob::learn::misc::KMeansTrainer::setAverageMinDistance, "Average min (square Euclidean) distance. Useful to parallelize the E-step.")
      .add_property("zeroeth_order_statistics", make_function(&bob::learn::misc::KMeansTrainer::getZeroethOrderStats, return_value_policy<copy_const_reference>()), &py_setZeroethOrderStats, "The zeroeth order statistics. Useful to parallelize the E-step.")
      .add_property("first_order_statistics", make_function(&bob::learn::misc::KMeansTrainer::getFirstOrderStats, return_value_policy<copy_const_reference>()), &py_setFirstOrderStats, "The first order statistics. Useful to parallelize the E-step.")
diff --git a/bob/learn/misc/old/main.cc b/bob/learn/misc/old/main.cc
index 119d403f14cc274140857c91030511e416338da6..593957ecf7272c385a217e1d59b98dde2b5f6a12 100644
--- a/bob/learn/misc/old/main.cc
+++ b/bob/learn/misc/old/main.cc
@@ -14,6 +14,7 @@
 #include <bob.blitz/capi.h>
 #include <bob.blitz/cleanup.h>
 #include <bob.io.base/api.h>
+#include <bob.core/random.h>
 
 #include "ndarray.h"
 
@@ -54,6 +55,12 @@ BOOST_PYTHON_MODULE(_old_library) {
     return;
   }
 
+  if (import_bob_core_random() < 0) {
+    PyErr_Print();
+    PyErr_Format(PyExc_ImportError, "cannot import `bob.core.random'");
+    return;
+  }
+
   if (import_bob_io_base() < 0) {
     PyErr_Print();
     PyErr_Format(PyExc_ImportError, "cannot import `bob.io.base'");
diff --git a/bob/learn/misc/old/plda_trainer.cc b/bob/learn/misc/old/plda_trainer.cc
index 62c39d9e81626b2fb79c5157c9141dd55141575b..1c31059c9b2fe64664b845d0a4baa26fcd29b6e2 100644
--- a/bob/learn/misc/old/plda_trainer.cc
+++ b/bob/learn/misc/old/plda_trainer.cc
@@ -92,11 +92,28 @@ static object get_z_second_order(bob::learn::misc::PLDATrainer& m) {
 }
 
 
+// include the random API of bob.core
+#include <bob.core/random.h>
+static boost::python::object TB_getRng(EMTrainerPLDA& self){
+  // create new object
+  PyObject* o = PyBoostMt19937_Type.tp_alloc(&PyBoostMt19937_Type,0);
+  reinterpret_cast<PyBoostMt19937Object*>(o)->rng = self.getRng().get();
+  return boost::python::object(boost::python::handle<>(o));
+}
+
+#include <boost/make_shared.hpp>
+static void TB_setRng(EMTrainerPLDA& self, boost::python::object rng){
+  if (!PyBoostMt19937_Check(rng.ptr())) PYTHON_ERROR(TypeError, "Would have expected a bob.core.random.mt19937 object");
+  PyBoostMt19937Object* o = reinterpret_cast<PyBoostMt19937Object*>(rng.ptr());
+  self.setRng(boost::make_shared<boost::mt19937>(*o->rng));
+}
+
+
 void bind_trainer_plda()
 {
   class_<EMTrainerPLDA, boost::noncopyable>("EMTrainerPLDA", "The base python class for all EM/PLDA-based trainers.", no_init)
     .add_property("max_iterations", &EMTrainerPLDA::getMaxIterations, &EMTrainerPLDA::setMaxIterations, "Max iterations")
-    .add_property("rng", &EMTrainerPLDA::getRng, &EMTrainerPLDA::setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
+    .add_property("rng", &TB_getRng, &TB_setRng, "The Mersenne Twister mt19937 random generator used for the initialization of subspaces/arrays before the EM loop.")
     .def("train", &plda_train, (arg("self"), arg("machine"), arg("data")), "Trains a PLDABase using data (mu, F, G and sigma are learnt).")
     .def("initialize", &plda_initialize, (arg("self"), arg("machine"), arg("data")), "This method is called before the EM algorithm")
     .def("finalize", &plda_finalize, (arg("self"), arg("machine"), arg("data")), "This method is called at the end of the EM algorithm")