diff --git a/index.js b/index.js
index 9a4ed19..bb0eaf4 100644
--- a/index.js
+++ b/index.js
@@ -1,8 +1,8 @@
 /*
  * This library is developed by the ISMM (http://ismm.ircam.fr/) team at IRCAM,
  * within the context of the RAPID-MIX (http://rapidmix.goldsmithsdigital.com/)
- * project, funded by the European Union’s Horizon 2020 research and innovation programme.  
- * Original XMM code authored by Jules Françoise, ported to Node.js by Joseph Larralde.  
+ * project, funded by the European Union’s Horizon 2020 research and innovation programme.
+ * Original XMM code authored by Jules Françoise, ported to Node.js by Joseph Larralde.
  * See https://github.com/Ircam-RnD/xmm for detailed XMM credits.
  */
 
@@ -17,9 +17,9 @@ var XmmNative = null;
 // Load the precompiled binary for windows.
 
 // if (process.platform == "win32" && process.arch == "x64") {
-//  XmmNative = require('./bin/winx64/xmm');  
+//  XmmNative = require('./bin/winx64/xmm');
 // } else if(process.platform == "win32" && process.arch == "ia32") {
-//  XmmNative = require('./bin/winx86/xmm');  
+//  XmmNative = require('./bin/winx86/xmm');
 // } else {
 
 var binary = require('node-pre-gyp');
@@ -67,6 +67,8 @@ function translateFromXmmConfigProp(prop) {
     return 'transitionMode';
   } else if (prop === 'regression_estimator') {
     return 'regressionEstimator';
+  } else if (prop === 'multiClass_regression_estimator') {
+    return 'multiClassRegressionEstimator';
   } else {
     return prop;
   }
@@ -83,6 +85,8 @@ function translateToXmmConfigProp(prop) {
     return 'transition_mode';
   } else if (prop === 'regressionEstimator') {
     return 'regression_estimator';
+  } else if (prop === 'multiClassRegressionEstimator') {
+    return 'multiClass_regression_estimator';
   } else {
     return prop;
   }
@@ -106,7 +110,7 @@ Xmm.prototype.getConfig = function(prop) {
 Xmm.prototype.setConfig = function(config) {
   var inConfig = {};
   for (var prop in config) {
-    var translatedProp = translateToXmmConfigProp(prop);    
+    var translatedProp = translateToXmmConfigProp(prop);
     inConfig[translatedProp] = config[prop];
   }
   this._xmm.setConfig(inConfig);
diff --git a/package.json b/package.json
index e592894..c89b7a2 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "xmm-node",
-  "version": "0.3.2",
+  "version": "0.3.3",
   "author": "Joseph Larralde",
   "license": "GPL-3.0",
   "description": "node wrapper for the XMM library",
@@ -42,7 +42,7 @@
     "jsdoc-template": "ircam-jstools/jsdoc-template",
     "jsdoc-to-markdown": "^1.3.7",
     "tape": "^4.6.0",
-    "xmm-client": "ircam-rnd/xmm-client"
+    "xmm-client": "github:ircam-rnd/xmm-client#v0.3.3"
   },
   "babel": {
     "presets": [
diff --git a/src/GmmTool.h b/src/GmmTool.h
index 8cdd551..989eaa8 100644
--- a/src/GmmTool.h
+++ b/src/GmmTool.h
@@ -9,8 +9,8 @@
 #include <iostream>
 #include "XmmTool.h"
 
-class GmmTool : public XmmTool<xmm::GMM> {
-    
+class GmmTool : public XmmTool<xmm::GMM, xmm::GMM> {
+
 public:
   GmmTool() {};
   ~GmmTool() {};
diff --git a/src/HhmmTool.h b/src/HhmmTool.h
index 662022f..252097b 100644
--- a/src/HhmmTool.h
+++ b/src/HhmmTool.h
@@ -9,12 +9,12 @@
 #include <iostream>
 #include "XmmTool.h"
 
-class HhmmTool : public XmmTool<xmm::HierarchicalHMM> {
-    
+class HhmmTool : public XmmTool<xmm::HierarchicalHMM, xmm::HMM> {
+
 public:
   HhmmTool() {};
   ~HhmmTool() {};
-    
+
   void setNbStates(std::size_t nbStates) {
     if(nbStates > 0) {
       model.configuration.states.set(nbStates, 1);
diff --git a/src/XmmTool.h b/src/XmmTool.h
index 39e022b..31c5f81 100644
--- a/src/XmmTool.h
+++ b/src/XmmTool.h
@@ -26,6 +26,8 @@ class XmmToolBase {
   virtual void setAbsoluteRegularization(double absReg) = 0;
   virtual xmm::GaussianDistribution::CovarianceMode getCovarianceMode() = 0;
   virtual void setCovarianceMode(xmm::GaussianDistribution::CovarianceMode cm) = 0;
+  virtual xmm::MultiClassRegressionEstimator getMultiClassRegressionEstimator() = 0;
+  virtual void setMultiClassRegressionEstimator(xmm::MultiClassRegressionEstimator mre) = 0;
 };
 
 // the specializable template :
@@ -36,7 +38,7 @@ class XmmTool : public XmmToolBase {
 private:
   std::vector<XmmWrapTrainWorker<Model> *> workers;
   // std::vector<Nan::Callback *> callbacks;
-  
+
 public:
   Model model;
 
@@ -49,7 +51,7 @@ class XmmTool : public XmmToolBase {
   }
 
   ~XmmTool() {}
-  
+
   void setBimodal(bool multimodality) {
     // Model tmp = Model(model);
     xmm::Configuration<ModelType> config = model.configuration;
@@ -101,7 +103,7 @@ class XmmTool : public XmmToolBase {
 
   v8::Local<v8::Object> filter(std::vector<float> observation) {
     v8::Local<v8::Object> outputResults = Nan::New<v8::Object>();
-  
+
     bool bimodal = model.shared_parameters->bimodal.get();
     unsigned int nmodels = model.size();
     unsigned int dimension = model.shared_parameters->dimension.get();
@@ -158,7 +160,7 @@ class XmmTool : public XmmToolBase {
 
     if (bimodal) {
       v8::Local<v8::Array> output_values = Nan::New<v8::Array>(dimension_output);
-      for (unsigned int i = 0; i < dimension_output; ++i) { 
+      for (unsigned int i = 0; i < dimension_output; ++i) {
         Nan::Set(output_values, i, Nan::New(res.output_values[i]));
       }
       outputResults->Set(
@@ -168,7 +170,7 @@ class XmmTool : public XmmToolBase {
 
       unsigned int dim_out_cov = res.output_covariance.size();
       v8::Local<v8::Array> output_covariance = Nan::New<v8::Array>(dim_out_cov);
-      for (unsigned int i = 0; i < dim_out_cov; ++i) { 
+      for (unsigned int i = 0; i < dim_out_cov; ++i) {
         Nan::Set(output_covariance, i, Nan::New(res.output_covariance[i]));
       }
       outputResults->Set(
@@ -218,6 +220,15 @@ class XmmTool : public XmmToolBase {
     model.configuration.changed = true;
   }
 
+  xmm::MultiClassRegressionEstimator getMultiClassRegressionEstimator() {
+    return model.configuration.multiClass_regression_estimator;
+  }
+
+  void setMultiClassRegressionEstimator(xmm::MultiClassRegressionEstimator mre) {
+    model.configuration.multiClass_regression_estimator = mre;
+    model.configuration.changed = true;
+  }
+
 };
 
 #endif /* _XMM_TOOL_H_ */
diff --git a/src/XmmWrap.cpp b/src/XmmWrap.cpp
index f8a55ae..8f786d5 100644
--- a/src/XmmWrap.cpp
+++ b/src/XmmWrap.cpp
@@ -456,6 +456,7 @@ void XmmWrap::getConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
   double relative_regularization = obj->model_->getRelativeRegularization();
   double absolute_regularization = obj->model_->getAbsoluteRegularization();
   xmm::GaussianDistribution::CovarianceMode covariance_mode = obj->model_->getCovarianceMode();
+  xmm::MultiClassRegressionEstimator multiclass_regression_estimator = obj->model_->getMultiClassRegressionEstimator();
 
   // HierarchicalHMM-specific :
   bool hierarchical = true;
@@ -490,6 +491,12 @@ void XmmWrap::getConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
                        ? "likeliest"
                        : ""));
 
+  std::string mre = (multiclass_regression_estimator == xmm::MultiClassRegressionEstimator::Likeliest)
+                  ? "likeliest"
+                  : ((multiclass_regression_estimator == xmm::MultiClassRegressionEstimator::Mixture)
+                    ? "mixture"
+                    : "");
+
   std::string modelType;
   switch (obj->modelType_) {
     case XmmGmmE:
@@ -522,6 +529,8 @@ void XmmWrap::getConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
       args.GetReturnValue().Set(Nan::New<v8::Number>(absolute_regularization));
     } else if (item == "covariance_mode") {
       args.GetReturnValue().Set(Nan::New<v8::String>(cm).ToLocalChecked());
+    } else if (item == "multiclass_regression_etimator") {
+      args.GetReturnValue().Set(Nan::New<v8::String>(mre).ToLocalChecked());
     }
 
     if (obj->modelType_ == XmmHhmmE) {
@@ -547,6 +556,8 @@ void XmmWrap::getConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
               Nan::New<v8::Number>(absolute_regularization));
     outputConfig->Set(Nan::New<v8::String>("covariance_mode").ToLocalChecked(),
               Nan::New<v8::String>(cm).ToLocalChecked());
+    outputConfig->Set(Nan::New<v8::String>("multiClass_regression_estimator").ToLocalChecked(),
+              Nan::New<v8::String>(mre).ToLocalChecked());
 
     if (obj->modelType_ == XmmHhmmE) {
       outputConfig->Set(Nan::New<v8::String>("hierarchical").ToLocalChecked(),
@@ -577,6 +588,7 @@ void XmmWrap::setConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
     double rr = obj->model_->getRelativeRegularization();
     double ar = obj->model_->getAbsoluteRegularization();
     xmm::GaussianDistribution::CovarianceMode cm = obj->model_->getCovarianceMode();
+    xmm::MultiClassRegressionEstimator mre = obj->model_->getMultiClassRegressionEstimator();
 
     // HierarchicalHMM-specific :
     bool h = true;
@@ -659,12 +671,25 @@ void XmmWrap::setConfig(const Nan::FunctionCallbackInfo<v8::Value> & args) {
       }
     }
 
+    v8::Local<v8::Value> multiclass_regression_estimator
+      = inputConfig->Get(Nan::New<v8::String>("multiClass_regression_estimator").ToLocalChecked());
+    if (multiclass_regression_estimator->IsString()) {
+      v8::String::Utf8Value val(multiclass_regression_estimator->ToString());
+      std::string smre = std::string(*val);
+      if (smre == "likeliest") {
+        mre = xmm::MultiClassRegressionEstimator::Likeliest;
+      } else if (smre == "mixture") {
+        mre = xmm::MultiClassRegressionEstimator::Mixture;
+      }
+    }
+
     // =============== SET NEW VALUES ============== //
 
     obj->model_->setGaussians(g);
     obj->model_->setRelativeRegularization(rr);
     obj->model_->setAbsoluteRegularization(ar);
     obj->model_->setCovarianceMode(cm);
+    obj->model_->setMultiClassRegressionEstimator(mre);
 
     if (obj->modelType_ == XmmHhmmE) {
       setHierarchical(obj, h);
diff --git a/test/1_basic_tests.js b/test/1_basic_tests.js
index 43abfe2..bd6612f 100644
--- a/test/1_basic_tests.js
+++ b/test/1_basic_tests.js
@@ -15,6 +15,7 @@ test('model configuration consistency', (t) => {
 		covarianceMode: 'full',
 		relativeRegularization: 0.1,
 		absoluteRegularization: 0.1,
+		multiClassRegressionEstimator: 'mixture',
 		// states: 1,
 	};
 
diff --git a/test/2_training_tests.js b/test/2_training_tests.js
index 62976f1..5dc241e 100644
--- a/test/2_training_tests.js
+++ b/test/2_training_tests.js
@@ -1,5 +1,5 @@
 import xmm from '../index';
-import { SetMaker } from 'xmm-client';
+import { GmmDecoder, SetMaker } from 'xmm-client';
 import test from 'tape';
 
 test('training', (t) => {
@@ -8,14 +8,17 @@ test('training', (t) => {
 
   const setMaker = new SetMaker();
 
-  const hhmm = new xmm('hhmm', {
-    hierarchical: false,
-    relativeRegularization: 0.1
+  const gmm = new xmm('gmm', {
+    // hierarchical: false,
+    relativeRegularization: 0.1,
+    multiClassRegressionEstimator: 'mixture',
   });
 
+  const gmmClient = new GmmDecoder();
+
   var p = {
     bimodal: true,
-    dimension: 3,
+    dimension: 6,
     dimension_input: 3,
     column_names: [ "" ],
     // data: [
@@ -24,39 +27,57 @@ test('training', (t) => {
     //   3.7, 3.2
     // ],
     data_input: [1, 2, 3],
-    data_output: [1, 2, 3],
+    data_output: [0, 0, 0],
     length: 1,
     label: 'aLabel'
-  }
+  };
 
   const trainMsgOne = 'train should return a null model when training is cancelled';
 
-  hhmm.train((err, res) => {
+  gmm.train((err, res) => {
     t.equal(res, null, trainMsgOne);
   });
 
-  hhmm.cancelTraining();
+  gmm.cancelTraining();
 
   const trainMsgTwo = 'train should return an empty model when trained with empty set';
 
-  hhmm.train((err, res) => {
+  gmm.train((err, res) => {
     t.deepEqual(res.models, [], trainMsgTwo);
   });
 
   const trainMsgThree = 'train should return a trained model';
 
-  for (let i = 0; i < 500; i++) {
-    setMaker.addPhrase(p);
+  for (let i = 0; i < 5; i++) {
+    setMaker.addPhrase(JSON.parse(JSON.stringify(p)));
     // hhmm.addPhrase(JSON.parse(JSON.stringify(p)));
   }
-  // hhmm.setTrainingSet(setMaker.getTrainingSet());
+
+  p.data_input = [3, 2, 1];
+  p.data_output = [10, 10, 10];
+  p.label = 'anotherLabel';
+
+  for (let i = 0; i < 5; i++) {
+    setMaker.addPhrase(JSON.parse(JSON.stringify(p)));
+    // hhmm.addPhrase(JSON.parse(JSON.stringify(p)));
+  }
+
+  const set = setMaker.getTrainingSet();
+  gmm.setTrainingSet(set);
+  console.log(set);
   // hhmm.clearTrainingSet();
   // hhmm.addTrainingSet(setMaker.getTrainingSet());
 
-  hhmm.train((err, res) => {
+  gmm.train((err, res) => {
+    console.log(JSON.stringify(res, null, 2));
+    gmmClient.setModel(res);
+
     t.notEqual(res, null, trainMsgThree);
     t.equal(res.models.length > 0, true, trainMsgThree);
 
+    console.log(gmmClient.filter([3, 2, 3]));
+    console.log(gmmClient.filter([3, 2, 1]));
+
     // const setModelConfigMsg = 'config should not change when queried after setModel';
     // let config = hhmm.getConfig();
     // hhmm.setModel(res);