Skip to main content

Table 4 Average CPU time breakdown over test set 2-7

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

  CPU time (ms/frame) BG calculated (%)
  Mask estimation BG Calculation CG Evaluation Beam Search Total  
1. No MDT 0.0 3.5 0.2a 24.4 28.1 27
2. BE PROSPECT MDT 1.1 133.5 0.0 18.5 153.1 100
3. GB GD MC MDT 1.1 2.8 4.3 20.0 28.2 11
4. GB MAP MC MDT 1.1 3.3 7.6 22.2 34.2 13
5. GB GD MC MDT+CGS 1.1 2.4 1.2 19.1 23.8 9
6. GD CLBR 1.1 1.3 4.3 19.3 26.0 11
7. MAP CLBR 1.1 1.2 10.6b 18.5 31.4 9
8. SB MC MDT 1.1 30.2 4.6 20.1 56.0 11
9. Spectral MC MDT 1.1 10.4 0.5 23.7 35.7 38
10.Cepstral MC MDT 1.1 2.8 7.0 19.0 29.9 10.5
  1. aIn the case of No MDT, the time for CG evaluation is replaced by FRoG. bIn cluster-based reconstruction with MAP, the time of CG evaluation includes the calculation of the likelihood for 500 PROSPECT CGs, which is equivalent to evaluation of 1000 MIDA Gaussians.
  2. GB, Gaussian-based; SB, State-based; CGS, CG selection; CLBR, Cluster-based reconstruction; BE, Backend.