Speech recognition engine has difficulties recognizing separate letters (nl_NL)

Vergeest, LC (Lucas) 1 Reputation point
2021-10-29T07:54:39.203+00:00

I have noticed that when pronouncing separate letters in a test set, the engine has a lot of difficulty to recognize them (Dutch language).
Of course this is not a big surprise, since there is no meaningful context, and the phoneme clusterse are relatively short.

However, I was wondering, does anyone have suggestions as how to improve the WER for this?

Perhaps just adding a training set of separate letters, including annotations?
Or would it be better to encourage customers to pronounce letters as full names, eg. when pronouncing zip codes or license plates?

Azure AI Speech
Azure AI Speech
An Azure service that integrates speech processing into apps and services.
1,774 questions
0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. GiftA-MSFT 11,171 Reputation points
    2021-11-01T21:39:44.047+00:00

    Hi, the first step to improve recognition is to identify the type of error (Insertion, Deletion , Substitution). You can improve recognition results by adding training data such as related text sentences, audio with human-labeled transcripts, new words with pronunciation. However, understanding the type of error that exists helps you to apply the best approach. Please review Evaluate and improve Custom Speech Accuracy for more details.

    --- *Kindly Accept Answer if the information helps. Thanks.*

    0 comments

Your answer

Answers can be marked as Accepted Answers by the question author, which helps users to know the answer solved the author's problem.