hmm_09.cc

这是一个从音频信号里提取特征参量的程序

// file: $isip/class/pr/HiddenMarkovModel/hmm_09.cc
// version: $Id: hmm_09.cc,v 1.1 2003/01/11 15:20:59 jelinek Exp $
//

// isip include files
//
#include "HiddenMarkovModel.h"

// method: adaptPart
//
// arguments:
//  Vector<MatrixDouble>& g: (output) vector of G matrices
//  MatrixFloat& z: (output) Z matrix
//  GaussianModel& gm: (input) Gaussian model with its statistics
//
// return: a boolean value indicating status
//
// this method computes contribution of one Gaussian model to
// cumulative adaptation matrices G and Z
//
boolean HiddenMarkovModel::adaptPart(Vector<MatrixDouble>& g_a,
				     MatrixFloat& z_a,
				     GaussianModel& gm_a) {

  // print the access count in debug mode
  //
  if (debug_level_d >= Integral::ALL) {
    String out(L"access count:");
    out.concat(gm_a.getAccessCount());
    Console::put(out);
  }

  // if model was accessed, compute its contribution to the global
  // adaptation and add it to the global adaptation matrices
  //
  if (gm_a.getAccessCount() > 0) {

    // form the extended mean
    //
    VectorFloat extended_mean;
    gm_a.getMean(extended_mean);

    long num_feat = extended_mean.length();
    extended_mean.setLength(num_feat + 1);
    for (long i = num_feat; i > 0; i--) {
      extended_mean(i) = extended_mean(i - 1);
    }
    extended_mean(0) = 1;

    // inverse the covariance matrix
    //
    MatrixFloat inv_cov;
    gm_a.getCovariance(inv_cov);
    inv_cov.inverse();
      
    // convert mean accumulator to float
    //
    VectorDouble mean_accum;
    gm_a.getMeanAccumulator(mean_accum);
    VectorFloat mean_accum_f(num_feat);
    for (long i = 0; i < num_feat; i++) {
      mean_accum_f(i) = (Float)mean_accum(i);
    }

    // compute contribution of this model to the matrix Z
    //
    VectorFloat tmp;
    inv_cov.multv(tmp, mean_accum_f);

    MatrixFloat z_r;
    z_r.outerProduct(tmp, extended_mean);
    z_a.add(z_r);

    // compute V matrix
    //
    MatrixFloat v(inv_cov);
    Double occup_accum = gm_a.getOccupancy();
    v.mult((float)occup_accum);

    // compute D matrix
    //
    MatrixFloat d;
    d.outerProduct(extended_mean, extended_mean);

    // check for the diagonal covariance before computing G matrix
    //
    if (!inv_cov.isDiagonal()) {
      return Error::handle(name(),
			   L"adapt - only diagonal matrices supported",
			   Error::ARG, __FILE__, __LINE__);
    }
    
    // compute contribution of this model to the matrix G
    //
    MatrixFloat g_r;
    for (long i = 0; i < num_feat; i++) {
      g_r.assign(d);
      g_r.mult(v(i, i));
      g_a(i).add(g_r);
    }

    // exit gracefully
    //
    return true;
  }

  // else the model was not accessed
  //
  else {
    
    // exit ungracefully
    //
    return false;
  }
}