custom STT: Enhancing the deletion error of custom model

Question

Hi,

I am the custom STT user training with korean, and want some training advice related to deletion errors.

The model I have trained gets better on insertion and substitution error, but due to huge decline in deletion error, the overall quality has worsen than the base model.

According to the doucment above, it said :

"When many deletion errors are encountered, it's usually because of weak audio signal strength. To resolve this issue, you need to collect audio data closer to the source."

I thought this was about testset audio signal strength, so I trained the model with small sound volume (20 db smaller that the original trainset) to increase its sensitivity to small sound. The

deletion error got improved somehow but it is still an major issue for the quality deterioration.

Is there any other way I can try to improve my deletion error rate? Any advice would be welcomed.

below is the training and test result compared to the base model for your information.

                          insertion           substition               deletion                 overall  
dataset1 (original)      -1.38%               -2.63%              **+5.23%**            +1.22% (worsen)  
dataset2  (-20db)       -1.35%              -2.61%               **+2.68%**            -1.28% (better)  

  

Answer 1

@ywmo Thanks for the question. we would recommend to raise a Azure support desk ticket from Help+Support blade from Azure portal for your resource if you have a support plan for your subscription. This will help you to share the details securely and work with an engineer who can provide more insights about the issue that if it can be replicated.