male, AM: GD triphone 2000x16
| accuracy | LM size |
| 20K 45month cutoff-1-1 | 89.8 | 54MB |
|
20K 45month compress10% | 89.3 | 28MB |
| 20K 75month cutoff-1-1 | 94.3 | 79MB |
| 20K 75month compress10% | 94.3 | 38MB |
| 60K 75month cutoff-1-1 | 93.7 | 100MB |
|
60K 75month compress10% | 93.5 | 55MB |
- Increase of training data leads to accuracy improvement
- ML-compression method does not bring error increase
- Slight degradation with 60K model
Next: Decoding Algorithm Evaluation
Up: Performance Evaluation
Previous: Acoustic Model Evaluation
Tatsuya Kawahara
5/31/2000