Skip to main content
Figure 2 | EURASIP Journal on Audio, Speech, and Music Processing

Figure 2

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

Figure 2

The process of calculating the likelihoods of the BGs in the Gaussian-based MC MDT system. For each of the M CG (left), a clean speech spectrum is imputed using MDT and solutions with high posterior probability (CG selection table) are transformed to the feature representation of the backend (imputed MIDA). BGs (right) are evaluated at the candidates listed in the association table (crosses) and the highest likelihood is retained (bottom).

Back to article page