Research Article
Convolutional Neural Networks to Facilitate the Continuous Recognition of Arabic Speech with Independent Speakers
Table 4
WER and accuracy results of Sphinx-4 recognizer.
| MFCC features | Gaussian 2 | Gaussian 4 | Gaussian 8 | Gaussian 16 | Gaussian 32 | WER (%) | Acc. (%) | WER (%) | Acc. (%) | WER (%) | Acc. (%) | WER (%) | Acc. (%) | WER (%) | Acc. (%) |
| 13 features | 13.39 | 86.61 | 11.02 | 88.98 | 11.13 | 88.87 | 10.85 | 89.15 | 14.25 | 85.75 | 25 features | 10.77 | 89.23 | 8.45 | 91.55 | 8.17 | 91.83 | 8.65 | 91.35 | 10.85 | 89.15 | 39 features | 14.70 | 85.30 | 12.78 | 87.22 | 10.91 | 89.09 | 10.77 | 89.23 | 13.86 | 86.14 | 52 features | 23.60 | 76.56 | 20.81 | 79.19 | 18.41 | 81.59 | 18.33 | 81.67 | 32.01 | 76.99 |
|
|
The bold values indicate the best results of Sphinx-4 recognizer from WER and accuracy.
|