Skip to main content

Table 1 The expensive operations in the different methods for Gaussian-based MLE imputation with full covariance matrices

From: Multi-candidate missing data imputation for robust speech recognition

  Full precision matrix in cache Number of iterations Likelihood #multiplications Step calculation #multiplications Other
MAP [24] Yes 6 (D + 1)D (D t, u + 1)D t, u No
MU [32] Yes 5 (D + 1)D D t, u D t, u + 4D t, u D t, u √ per step
GD + cepstral No 2 2D m D + D m + D D t, u (2D m + 3) + D m D + D m + D No
GD + PROSPECT No 2 2(D c + D) 3D c D t, u + 4D c D + 5D + 3D c No
  1. D is the dimension of the log spectral features. D m is the order of the cepstral coefficients in MFCC. D t, u is the number of unreliable components at frame t. D c is the order of the cepstra in PROSPECT features. Typical values: D = 22, D m = 13, D c = 4. D t, u is about 16 on average as measured in [33].