Pronunciation Score not consistent for domain specific terms.

Question

I am using pronunciation assessment for domain specific terms but pron score is not consistent or low while tested via the speech endpoint(on baseline model). I have also trained a custom speech model using multiple audio terms of 2-3 seconds duration and tested it using that endpoint, but the scores are still inaccurate on lower side(below 40 out of 100).

Answer 1

Hello Sagar Kulkarni

There are several ways to improve pronunciation assessment scores in Azure. Here are some tips that you can try:

1. Use high-quality audio: Make sure that the audio you use for testing is of high quality and free from background noise.
2. Use domain-specific models: If you’re using domain-specific terms, consider training a custom speech model that is specific to your domain.
3. Use streaming mode: Pronunciation assessment supports uninterrupted streaming mode, which allows you to record audio for an unlimited amount of time.
4. Use a microphone: Make sure that you’re using a high-quality microphone when recording audio.
5. Speak clearly and slowly: When recording audio, speak clearly and slowly to ensure that your pronunciation is accurate.

You can find more information about pronunciation assessment in Microsoft’s official documentation.

Please don’t forget to Accept Answer and Yes for "was this answer helpful" wherever the information provided helps you, this can be beneficial to other community members.