Looking to use specific speech-to-text base model: How long would the base model exist for?

Question

Looking to use specific speech-to-text base model: How long would the base model exist for?

Caesar Cavales 50

Our team is using Azure speech-to-text in our production applications and I have notice a difference in result from time to time when we try to run a bulk transcription to test accuracy. When I investigated, I notice that we might be calling the most recent base model when we did our testing.

We have decided to use/pin the same base model until we have tested and ran benchmarks on any new base model that becomes available. Before I make the change to pin the same base model whenever we call the REST API, I am curious for how long a model would be available for. The last thing we want to happen is that we would get back an error because a base model we are using, hardcoded, in our code does not exist anymore.

Accepted answer

0 additional answers

Your answer

Answer 1

romungi-MSFT 48,906 Microsoft Employee Moderator

@Caesar Cavales You can check the base model deprecation dates using the GET Base model API. This provides a deprecation date for the base model that you can use to determine when a base model is going to expire. I think you can use this API to automate your application deployments if the base model date is valid.

Since there are many models, use the filter expression to use your base model display name. For example the request URL should look like something below along with subscription key header.

https://<resource_name>.cognitiveservices.azure.com/speechtotext/v3.1/models/base?filter=displayName eq '20221013'

Query Filter:

filter=displayName eq '20221013'

Sample response:

User's image

If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.

Caesar Cavales 50 Reputation points

2023-07-25T03:21:27.9433333+00:00
Hi @romungi-MSFT , thank you for your reply. I think this is something I can work with. Just a couple followup questions.

In the "deprecationDates" object, which key should I be looking at? There is "adaptationDateTime" and "transcriptionDateTime".

For US-based apps, I should only be using models that have locale='en-US'?
romungi-MSFT 48,906 Reputation points Microsoft Employee Moderator

2023-07-25T06:40:27.51+00:00

Caesar Cavales The adaption date indicates that you cannot create a new endpoint or deployment after this date, all existing deployments will continue to work if they have been deployed before this adaption date till the transcription date time. So, your existing deployment should be available till the end of transcription date since they are already in use.

Yes, depending on locales choose the appropriate model.
Caesar Cavales 50 Reputation points

2023-07-25T15:58:26.0766667+00:00

In the case where we are calling the REST endpoint and that no deployments are made, we should be able to call any base model until the "transcriptionDateTime", correct?
romungi-MSFT 48,906 Reputation points Microsoft Employee Moderator

2023-07-26T05:05:09.18+00:00

If you have not used a base model for deployment, then you will have to check the "adaptationDateTime" and use the model for deployment before this date. Then you can use the "transcriptionDateTime" as end of life for the model. Thanks!!
Caesar Cavales 50 Reputation points

2023-07-26T15:23:00.1333333+00:00
I'm a little bit confused. So in the case where I have a speech service resource in Azure, calling the STT endpoint would be this:

https://<<resource>>/stt/speech/recognition/conversation/cognitiveservices/v1?language=en-US

If I want to pin a base model to the request to make sure I would get a consistent result every time, it would look like this for example:

https://<<resource>>.cognitiveservices.azure.com/speechtotext/v3.1/models/base/3437cefb-480a-4e45-b5f1-ff607b83ae4f)

If I understood correctly, the "models/base/<<model_id>>" is an id taken from the GET base models API and has an expiration date of "transcriptionDateTime". Using the REST API this way does not require me to deploy any model(correct me if I am wrong but when using STT REST API out of the bat, I did not deploy any model).

With the information above, should I be looking at the "transcriptionDateTime" to determine if the model will still be available in Azure?
Caesar Cavales 50 Reputation points

2023-07-26T15:30:31.61+00:00

duplicate of above
romungi-MSFT 48,906 Reputation points Microsoft Employee Moderator

2023-07-27T10:44:18.06+00:00

Calling the short audio endpoint would always be the following.

https://<REGION_IDENTIFIER>.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1

If you want to use a specific base model you will have to deploy it as a custom model and use the cid query parameter to use it.

Base models can only be used directly using batch transcription and model endpoints like /models/{id} can be used for model management i.e create, delete, list, get model operations.
Caesar Cavales 50 Reputation points

2023-08-08T17:28:08.43+00:00

Thank you for your help and patience in answering my quetsions @romungi-MSFT

Share via

Looking to use specific speech-to-text base model: How long would the base model exist for?

0 additional answers

Your answer