Custom Speech quality issue.

nattawutb 1 Reputation point
2020-09-23T04:30:55.233+00:00

I am developing a speech to text solution using Azure Custom Speech API as a backend. I noticed the drop in the quality of custom trained models.

Previously, I was kind of impressed by the result of a custom trained model where I can train with Text data or Text associated with voice prompt.
The previous training usually took 45-60 minutes and the result is very good the model can pick up words with high accuracy.

However, recently I found that it took less than 10 minutes to train a model and the quality of the trained model is not there.
So, Could you please advise how to revert back to the previous algorithm? I would rather not trade off on training time vs model quality.

Best Regards,

Nattawut B

Azure AI Speech
Azure AI Speech
An Azure service that integrates speech processing into apps and services.
1,391 questions
0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. romungi-MSFT 41,866 Reputation points Microsoft Employee
    2020-09-23T10:24:59.37+00:00

    @nattawutb In the current scenario you can use a different algorithm or baseline model to create a new model with your training data. Here is a screen shot of the available baseline models

    26983-image.png

    We hope one of the baseline models available in this list was used for the older model so a new model can be created with similar criteria.