kaldi cmvn

cur_sum.AddRowSumMat(1.0, input_part , 0.0);  // input 列求和得到cur_sum

cur_sumsq.AddDiagMat2(1.0, input_part, kTrans, 0.0);  //Add the diagonal of a matrix times itself  *this = diag(M^T M) +  beta * *this (if trans == kTrans).

variance.Scale(1.0 / window_frames);   // E(x**2)

variance.AddVec2(-1.0 / (window_frames * window_frames), cur_sum);   // E(x**2)-  (E(x) /n)**2 得到方差

                  Add vector : *this = *this + alpha * rv^2  [element-wise squaring].

variance.ApplyPow(-0.5); // get inverse standard deviation.    标准差的倒数

output_frame.MulElements(variance);    (x-E(x))/delta 得到 标准差归一化

 

Vector<double> variance(cur_sumsq); kaldi向量

variance.Scale(1.0 / window_frames);   乘系数

variance.ApplyPow(-0.5); // get inverse standard deviation.  // 幂运算

 

double Plda::TransformIvector(const PldaConfig &config,
                              const VectorBase<double> &ivector,
                              int32 num_examples,
                              VectorBase<double> *transformed_ivector) const {
  double normalization_factor;
  transformed_ivector->CopyFromVec(offset_);
  transformed_ivector->AddMatVec(1.0, transform_, kNoTrans, ivector, 1.0);
  if (config.simple_length_norm)
    normalization_factor = sqrt(transformed_ivector->Dim())
      / transformed_ivector->Norm(2.0);
  else
    normalization_factor = GetNormalizationFactor(*transformed_ivector,
                                                  num_examples);
  if (config.normalize_length)
    transformed_ivector->Scale(normalization_factor);
  return normalization_factor;