SuperVectors.cpp

Go to the documentation of this file.

/*
This file is part of LIA_RAL which is a set of software based on ALIZE
toolkit for speaker recognition. ALIZE toolkit is required to use LIA_RAL.

LIA_RAL project is a development project was initiated by the computer
science laboratory of Avignon / France (Laboratoire Informatique d'Avignon -
LIA) [http://lia.univ-avignon.fr <http://lia.univ-avignon.fr/>]. Then it
was supported by two national projects of the French Research Ministry:
        - TECHNOLANGUE program [http://www.technolangue.net]
        - MISTRAL program [http://mistral.univ-avignon.fr]

LIA_RAL is free software: you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as
published by the Free Software Foundation, either version 3 of
the License, or any later version.

LIA_RAL is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public
License along with LIA_RAL.
If not, see [http://www.gnu.org/licenses/].

The LIA team as well as the LIA_RAL project team wants to highlight the
limits of voice authentication in a forensic context.
The "Person Authentification by Voice: A Need of Caution" paper
proposes a good overview of this point (cf. "Person
Authentification by Voice: A Need of Caution", Bonastre J.F.,
Bimbot F., Boe L.J., Campbell J.P., Douglas D.A., Magrin-
chagnolleau I., Eurospeech 2003, Genova].
The conclusion of the paper of the paper is proposed bellow:
[Currently, it is not possible to completely determine whether the
similarity between two recordings is due to the speaker or to other
factors, especially when: (a) the speaker does not cooperate, (b) there
is no control over recording equipment, (c) recording conditions are not
known, (d) one does not know whether the voice was disguised and, to a
lesser extent, (e) the linguistic content of the message is not
controlled. Caution and judgment must be exercised when applying speaker
recognition techniques, whether human or automatic, to account for these
uncontrolled factors. Under more constrained or calibrated situations,
or as an aid for investigative purposes, judicious application of these
techniques may be suitable, provided they are not considered as infallible.
At the present time, there is no scientific process that enables one to
uniquely characterize a persones voice or to identify with absolute
certainty an individual from his or her voice.]

Copyright (C) 2004-2010
Laboratoire d'informatique d'Avignon [http://lia.univ-avignon.fr]
LIA_RAL admin [alize@univ-avignon.fr]
Jean-Francois Bonastre [jean-francois.bonastre@univ-avignon.fr]
*/

#if !defined(ALIZE_SuperVectors_cpp)
#define ALIZE_SuperVectors_cpp

#include "SuperVectors.h"
#include<iostream>
#include<fstream>
#include<cstdio>
#include<cassert>
#include<cmath>

using namespace alize;
using namespace std;

// comments in .h

void modelToSv(const MixtureGD &M,RealVector <double> &v){
        unsigned long modelSize=M.getDistribCount();
        unsigned long vectSize=M.getVectSize(); 
        v.setSize(modelSize*vectSize);
        for (unsigned long i=0;i<modelSize;i++)
                for (unsigned long j=0;j<vectSize;j++)
                        v[i*vectSize+j]=M.getDistrib(i).getMean(j);
}

void svToModel(RealVector  <double> &v, MixtureGD &M) {
        unsigned long modelSize=M.getDistribCount();
        unsigned long vectSize=M.getVectSize(); 
        for (unsigned long i=0;i<modelSize;i++)
                for (unsigned long j=0;j<vectSize;j++)
                        M.getDistrib(i).setMean(v[i*vectSize+j],j);
}

/*void projectOnSubSpace(RealVector<double> &x, Matrix<double> &U, RealVector <double>&proj) {
        // trick is to transform x'=S'Sx in x'=(U(U'x))' so that no need to compute U explicitely
        Matrix<double> Utmp=U;
        if (U.rows() > U.cols())
                Utmp.transpose();       
        Matrix<double> xt=((DoubleMatrix)x).transpose();
        Matrix<double> spkFactors=Utmp*xt;              
        if (verboseLevel > 1) {
                cout <<  "(SuperVectors) SpkFactors: (";
                for (unsigned long i=0;i<spkFactors.rows();i++)
                        cout <<i<<":"<<spkFactors(i,0)<<",";                    
        }
        if (verboseLevel > 2) cout << "(SuperVectors) SpkFactors: "<<spkFactors.rows() <<" " <<spkFactors.cols() << "...";
        Matrix<double> Ut=Utmp.transpose();
        Matrix<double>  offset=Ut*spkFactors;
        if (verboseLevel > 2) cout << "(SuperVectors) offset: "<<offset.rows() <<" " <<offset.cols() <<"...";
        for (unsigned long i=0;i<offset.rows();i++) {
                proj[i]=offset(i,0);
        }
}*/

void projectOnSubSpace(RealVector<double> &x, Matrix<double> &U, RealVector <double>&proj) { //fast version
        // trick is to transform x'=S'Sx in x'=(U(U'x))' so that no need to compute U explicitely
        double *_u=U.getArray();
        RealVector <double> tmp;
        tmp.setSize(U.rows());
        tmp.setAllValues(0.0);
        proj.setAllValues(0.0);
        if (verboseLevel > 1) cout<<"x: ("<<x.size()<<") - U; ("<<U.rows()<<","<<U.cols()<<")"<<endl;
        for (unsigned long i=0;i<U.rows();i++) 
                for (unsigned long j=0;j<U.cols();j++) 
                        tmp[i]+=_u[i*U.cols()+j]*x[j];
        U.transpose();
        _u=U.getArray();        
        for (unsigned long i=0;i<U.rows();i++) 
                for (unsigned long j=0;j<U.cols();j++) 
                        proj[i]+=_u[i*U.cols()+j]*tmp[j];               
        U.transpose();
}


void computeNap(MixtureGD &M,Matrix <double> &U) {
        unsigned long svSize=M.getDistribCount()*M.getVectSize();       
        RealVector <double> v(svSize,svSize);
        v.setAllValues(0.0);
        RealVector <double> proj(svSize,svSize);                
        proj.setAllValues(0.0); 
        modelToSv(M,v);
        projectOnSubSpace(v,U,proj);
        v-=proj;
        svToModel(v,M);
}


// Estimate NAP channel effect with a fixed client supervector and put it into model
// 1. Project client vector on test vector
// 2. Project result on channel subspace
// 3. Compute NAP on client model
// 4. Add projection to client NAPed model
// Three same function are coded but none of them work right now!
void computeNAPChannelEffect(MixtureGD &T,MixtureGD &M,Matrix<double>&U) {
        unsigned long svSize=max(U.rows(),U.cols());
        RealVector <double> t(svSize,svSize);
        RealVector <double> m(svSize,svSize);
        RealVector <double> proj(svSize,svSize);
        RealVector <double> tmp(svSize,svSize); 
        RealVector <double> channel(svSize,svSize);     
        modelToSv(T,t);
        modelToSv(M,m); 
        double normT=0.0;
        double angle=0.0;
        for (unsigned long i=0;i<t.size();i++) // norm du test
                normT+=t[i]*t[i];
        for (unsigned long i=0;i<t.size();i++) // produit scalaire
                angle+=t[i]*m[i];
        if (verboseLevel > 1) cout << "# <t,m>/||t||2=" << angle/normT << endl;
        for (unsigned long i=0;i<t.size();i++) // norm du test
                proj[i]=(angle/normT)*t[i];     
        projectOnSubSpace(m,U,tmp); // Channel effect for test
        RealVector <double> mNap=m;
        mNap-=tmp;      
        projectOnSubSpace(proj,U,channel);
        channel+=mNap;
        if (debug) {
                double n=0.0;
                double a=0.0;   
                for (unsigned long i=0;i<t.size();i++) // norm du test
                        n+=channel[i]*channel[i];
                for (unsigned long i=0;i<t.size();i++) // produit scalaire
                        a+=t[i]*channel[i];
                cout << "# <model,test>/||model||2=" << a/n << endl;    
        }
        svToModel(channel,M);
}

// Get an estimate of channel effect for a model. 
//The image of projection of the SV client on the test vector is computed and the channel effect 
// estimate is done on this result
// This is computed with the Thales theorem
/*void computeNAPChannelEffect(MixtureGD &T,MixtureGD &M,Matrix<double>&U) {
        unsigned long svSize=max(U.rows(),U.cols());
        RealVector <double> t(svSize,svSize);
        RealVector <double> m(svSize,svSize);
        RealVector <double> proj(svSize,svSize);
        RealVector <double> tmp(svSize,svSize);
        RealVector <double> channel(svSize,svSize);     

        modelToSv(T,t);
        modelToSv(M,m); 
        projectOnSubSpace(t,U,proj); // Channel effect for test
        RealVector <double> tNap=t;
        tNap-=proj;
        projectOnSubSpace(m,U,tmp); // Channel effect for test
        RealVector <double> mNap=m;
        mNap-=tmp;

        double normT=0.0;
        double normM=0.0;
        for (unsigned long i=0;i<t.size();i++) // norm du test
                normT+=tNap[i]*tNap[i];
        for (unsigned long i=0;i<m.size();i++) // produit scalaire
                normM+=mNap[i]*mNap[i];
        cout << "# Expanding factor [" << sqrt(normM/normT) <<"]"<< endl;
        for (unsigned long i=0;i<channel.size();i++)
                channel[i]=sqrt(normM/normT)*proj[i];   

        channel+=mNap;
        double n=0.0;
        double a=0.0;   
        for (unsigned long i=0;i<t.size();i++) // norm du test
                n+=channel[i]*channel[i];
        for (unsigned long i=0;i<t.size();i++) // produit scalaire
                a+=t[i]*channel[i];
        cout << "# <model,test>/||model||2=" << a/n << endl;            
        svToModel(channel,M);
}*/

// Add UUt Mt to the napped SV for each model (same channel for every model)
/*void computeNAPChannelEffect(MixtureGD &T,MixtureGD &M,Matrix<double>&U) {
        unsigned long svSize=max(U.rows(),U.cols());
        RealVector <double> t(svSize,svSize);
        RealVector <double> m(svSize,svSize);
        RealVector <double> channel(svSize,svSize);
        modelToSv(T,t);
        projectOnSubSpace(t,U,channel); 
        computeNap(M,U);        
        modelToSv(M,m);         
        channel+=m;
        svToModel(channel,M);
}*/

//------------------------------------------------------------------------
// Weight part of fisher kernel -- Nicolas SCHEFFER (deprecated)
void getFisherWeightVector(const MixtureGD& world, const MixtureGD& clientMixture, RealVector<double> & v,Config& config) {
  double w,c;
  for (unsigned long i=0;i<world.getDistribCount();i++) {
    w=world.weight(i);
    c=clientMixture.weight(i);
    //v[i]=c/sqrt(w);
        v[i]=c/w;
  }
}


//------------------------------------------------------------------------
// Weight part of fisher kernel -- Nicolas SCHEFFER (deprecated)
void getKLVector(MixtureGD& model, RealVector<double> & v,Config& config) {
        unsigned long mSize=model.getDistribCount();
        unsigned long vSize=model.getVectSize();
        for (unsigned long i=0;i<mSize;i++) {
                DistribGD & d=model.getDistrib(i);
                double w=model.weight(i);
                for (unsigned long j=0;j<vSize;j++) {
                        //v[i*vSize+j]=d.getMean(j)*(sqrt(w*d.getCovInv(j)));
                        v[i*vSize+j]=d.getMean(j)*(sqrt(w*d.getCovInv(j)));
                }
        }
}

void getSuperVector(RealVector<double> &v,MixtureGD &aprioriModel,MixtureGD &clientMixture,Config &config) {
        if (config.getParam("superVector")=="SVMUBM"){ //weight supervector
                v.setSize(clientMixture.getDistribCount(),true);
                getFisherWeightVector(aprioriModel,clientMixture,v,config); // For an SVM/UBM system (have to force client Mixture to be ML estimate MAPConst, alpha=1);
            }
        else if (config.getParam("superVector")=="KL") {//kl supervector
                v.setSize(clientMixture.getDistribCount()*clientMixture.getVectSize(),true);            
                getKLVector(clientMixture,v,config);
        }
        else throw Exception("Cannot find supervector mode [kl|svmubm]",__FILE__,__LINE__);
}

#endif