Azure Speech to Text Grammar

Question

Azure Speech to Text Grammar

margotmg 1

Hi,
I've created a customized Speech Model in Azure Speech to text. I've conducted some experiments and in some cases, I get better results when using the grammar without the customized model than when using the model itself. How is this possible? How does the grammar exactly work? I assumed that a customized model would always get better results than any grammar (at least for the speaker it was designed to be used with)
Thanks!

1 answer

Your answer

Answer 1

YutongTie-MSFT 53,966 Moderator

Hello,

Thanks for reaching out to us. Basically the Azure Speech Service model is built by general data, which may cover more cases then customized model. In general cases it works better. Customized model sometimes has over-fit issue which may effect the result. It all depends on the specific scenario and also your customized model. If you can provide more details. We may find out more possible reasons.

Regards,
Yutong

margotmg 1 Reputation point

2021-04-08T15:50:28.977+00:00

Hi,
I have used a grammar with no customized model, and the recognizer would still recognize words that are out of the grammar. However, when using a customized model, this same grammar worked as a finite state grammar; the recognizer would not recognize any word outside the grammar. So, in order to see how grammar affects the recognition accuracy, I created sort of a "customized" model without any data (it was a base model which was not trained with any data, but worked as a customized model). I observed that the recognition accuracy was better when using a finite state grammar than when using an acoustically customized model.
However, when using a grammar that was NOT FINITE with a customized model, this would sometimes worsen the results.

With speaker 1 the best results were obtained with customized model. Worsen when using any grammar.
With speaker 2 the best results were obtained with a finite state grammar (with no acoustic customization).
Does this make any sense?
YutongTie-MSFT 53,966 Reputation points Moderator

2021-04-09T08:04:41.293+00:00

Thanks for the details, could you please send us the sample you are using so that we can take a look? If the sample is not confidential could you please send it to us by email at Azcommunity@microsoft.com with url of this thread?

It's a little difficult to investigate this without any sample/ testing data.

Appreciate your help.

Regards,
Yutong
margotmg 1 Reputation point

2021-04-09T11:48:18.24+00:00

@YutongTie-5848
Sure! What data do you need exactly? the endpoint_id, the subscription key and the region? or do you need anything else?
YutongTie-MSFT 53,966 Reputation points Moderator

2021-04-09T18:03:12.157+00:00

Thank you! You covered almost everything, but we need the sample testing input as well. ^^

Regards,
Yutong
margotmg 1 Reputation point

2021-04-10T17:20:17.023+00:00

@YutongTie-5848
Email sent! Thank you!

Margot
YutongTie-MSFT 53,966 Reputation points Moderator

2021-05-13T09:01:17.57+00:00

Update for this issue here. Thanks for your details and information to help us investigate. I have forwarded this issue to engineering team but there is no ETA still for when this will be optimized. But I will still track on my end. ^^

Regards,
Yutong

Share via

Azure Speech to Text Grammar

1 answer

Your answer