Azure Speech to Text: Use a different base model when posting a transcription request

Question

Azure Speech to Text: Use a different base model when posting a transcription request

Mehdi Aman 20

Hello dear community,

We use Azure AI Speech to services to transcribe audio files. We mainly use it to transcribe Moroccan dialect "ar-MA".

There was a recent update of the base model in "V3.0" that surprisingly degraded (to my sense) the accuracy of the transcription.

Since Moroccan use some French words while they speak, the previous base models used to accurately capture this nuance and transcribed French words in Latin. Here is an example:

Previous base model (much more accurate):

أيه قلتي ليا بغيتي تشوف واحد الإقتطاعات في

compte

ديالك شنو مكتوب على الإقتطاعات؟

Current base model :

ايوا قلتيلي ابغيتي اشوف واحد الاقطاعات في ديالك شنو مكتوب على الاقتطاعات

So is there anyway to get back to the previous base model (one of the previous four models actually) ?

Thank you,

Mehdi

dupammi 8,615 Reputation points Microsoft External Staff

2023-12-26T12:55:03.67+00:00

Hi @Mehdi Aman ,

Thank you for using the Microsoft Q&A.

Unfortunately, it's not possible to get back to the previous base model using the Azure Speech to Text. The base models are managed by Microsoft and updates are automatically applied to improve the accuracy of the transcription service. Microsoft Azure is continuously working to improve the accuracy levels of these services.

You can try creating a custom model using the Custom Speech Service to improve the accuracy of the transcription.

Here's the link to the custom speech services.

I hope you understand. Thank you!
Mehdi Aman 20 Reputation points

2023-12-26T13:00:12.7866667+00:00

Thanks @dupammi , That is unfortunate. I think it is time we move towards building a Custom Speech Model. The French/Arabic Chars mix was a nice thing though. If you can escalate it as user feedback. Cheers.

Accepted answer

0 additional answers

Your answer

dupammi 8,615 Reputation points Microsoft External Staff

2023-12-26T12:55:03.67+00:00

Hi @Mehdi Aman ,

Thank you for using the Microsoft Q&A.

Unfortunately, it's not possible to get back to the previous base model using the Azure Speech to Text. The base models are managed by Microsoft and updates are automatically applied to improve the accuracy of the transcription service. Microsoft Azure is continuously working to improve the accuracy levels of these services.

You can try creating a custom model using the Custom Speech Service to improve the accuracy of the transcription.

Here's the link to the custom speech services.

I hope you understand. Thank you!
Mehdi Aman 20 Reputation points

2023-12-26T13:00:12.7866667+00:00

Thanks @dupammi , That is unfortunate. I think it is time we move towards building a Custom Speech Model. The French/Arabic Chars mix was a nice thing though. If you can escalate it as user feedback. Cheers.

Answer 1

Hi @Mehdi Aman ,

Thank you for your understanding.

Sure, we will talk to our internal team regarding your concerned issue that it is not feasible to get back to the previous base model using the Azure Speech to Text.

However, we appreciate your feedback. I encourage you to submit a feedback through the Azure Feedback Forum. Here's the link to the Azure Feedback Forum: Post idea · Community (azure.com)

Hope you understand. Thank you!

If this answers your query, do click Accept Answer and Yes for was this answer helpful.

Share via

Azure Speech to Text: Use a different base model when posting a transcription request

0 additional answers

Your answer