sphinx-train ( 1.0.0.7 vs. 1.0.0.9)

[hiyassat]

i would like to share with you this experiment
on Sphinx-train
comparison between sphinx-train 1.0.0.7 and sphinx-train 1.0.0.9 ( Long audio
training )
corpus description
Arabic dictionary
•
Arabic phone list
•
training (40 minutes)
•
testing decoding (40 minutes)
•
same files use in both training
•
calculation and result :
1) timing:
sphinx-train 1.0.0.7
Training time 5 minutes
Decoding time 4.42 minutes

sphinx-train 1.0.0.9 ( Long audio
training )

Training time 23 minutes
Decoding time 1.36 minutes
results
sphinx-train 1.0.0.7 sphinx-train 1.0.0.9 ( Long audio training )
TOTAL Words 2469 2469
Correct 467 460
Errors 2082 2090
TOTAL Percent correct 18.91% 18.63%
Error % 84.33% 84.65%
Accuracy % 15.67% 15.35%
SENTENCE ERROR 95.0% (572/602) 96.3%
WORD ERROR RATE 84.3% (2082/2469) 84.7% (2082/2469)

with linguistic question
sphinx-train 1.0.0.7 sphinx-train 1.0.0.9
TOTAL Words 2469 2469
Correct 550 539
Errors 2011 2021
TOTAL Percent correct 22.28% 21.83%
Error % 81.45% 81.86%
Accuracy % 18.55% 18.14%
SENTENCE ERROR 95.7% (576/602) 96.7% (572/602)
WORD ERROR RATE 81.5% (2011/2469) 81.9% (2082/2469)

after adaptation :
sphinx-train 1.0.0.7 sphinx-train 1.0.0.9 ( Long audio
training )
TOTAL Words 2469 we cant do
Correct 763 we cant do
Errors 1824 we cant do
TOTAL Percent correct 30.90% we cant do
Error % 73.88% we cant do
Accuracy % 26.12% we cant do
SENTENCE ERROR 91.7% (552/602) we cant do
WORD ERROR RATE 73.9% (1824/2469) we cant do
improve in performance with adaptation = 40.81%
•
improve in performance when add adaptation and linguistic question = 66.69%
•

[Nickolay V. Shmyrev]

Thanks for sharing the results!