Table 3 WER (%) and the number of parameters of triphone forest DTAM systems on the 1992 WSJ non-verbalized 5K closed-test set
From: Decision tree-based acoustic models for speech recognition
System | Number of trees | % WER | Number of parameters |
|---|---|---|---|
Non-forest (MFCC) | 1 | 12.9 | 766k |
Non-forest with gender information (MFCC) | 1 | 11.9 | 770k |
Non-forest (MCMS) | 1 | 13.3 | 798k |
Non-forest (MCMS + MFCC concatenated) | 1 | 12.5 | 707k |
MCMS + MFCC | 2 | 10.7 | 1500k |
Acoustic partitioning (MFCC) | 4 | 10.9 | 747k |
Speaker clustering (MFCC) | 4 | 11.9 | 806k |