Azure Speech Transcriber Availability
# support
d
The changelog mentions that support for Azure Speech Transcriber was added, but this does not yet appear to be reflected in the available transcribers on your dashboard or when creating assistants via your API (https://api.vapi.ai/api#/Assistants/AssistantController_create). Do you have an estimate for when it will be available? Additionally, once it is, will we be able to select custom transcribers from Azure, such as a fine-tuned version?
v
@Diogo Gonçalves I'm glad to hear about your interest in the new Azure Speech Transcriber support! It's possible that the dashboard and API documentation may not have been updated yet to reflect the latest changes. However, since the changelog has mentioned the addition, it should be available soon. I would recommend checking back periodically or subscribing to our updates for the latest information. Regarding the selection of custom transcribers from Azure, such as a fine-tuned version, this feature typically depends on the level of integration between VAPI and Azure. If Azure's API supports custom models and VAPI integrates that functionality, then it should be possible. However, for the most accurate and current information, please reach out to our support team. They will be able to provide you with the latest details on the availability and usage of custom Azure transcribers. If you need any further assistance or have more questions, feel free to ask!
You can continue this conversation by mentioning me in the message.
d
I would like to have an ETA for the availability, and also to know whether we will be able to select fine-tuned versions. I ask vapi team to reply to this question
b
Hi @Diogo Gonçalves it should be available shortly. I'm not familiar with the implementation details but i'm assuming BYOK should allow for custom transcribers?
d
Thanks for the answer. Could you provide a date?
b
Hopefully, in a few days it will be available in the dashboard and API. I believe you can still create it through the API with provider: 'azure'
d
b
@Diogo Gonçalves I understand! It should be in the docs within a few days
d
Thanks. Please keep me updated!
b
Hi @Diogo Gonçalves , wanted to update you on the progress. It's available in the dashboard now
d
Hi Bryant! It's only available in English. Will it be available in other languages? My use case is in Portuguese (PT-PT). Also, are you planning to provide the option to select fine-tuned model deployments?
The ability to select different languages may not be a requirement for me in fact as long as we can provide a custom endpoint or resource - that would allow me to point to a fine-tuned model deployment in Portuguese.
b
@Diogo Gonçalves Perhaps try BYOK or the custom transcriber option? That may fit your use case
RE: Other languages, I believe that's something we may add in the future, will have to check
d
With BYOK, the only available transcription provider is Deepgram (https://docs.vapi.ai/customization/provider-keys). Azure doesn’t appear to be supported, correct? Using a custom transcriber, the latency is quite high (over 1 second for transcription - I suspect it might be consequence of relying only on final results), which is why we were hoping to leverage your integration. Having the option to select a fine-tuned model deployment would be ideal—could you check if that's possible?
b
it's available under Transcriber Providers. it seems the docs for transcriber BYOK are not up to date
d
Ok, we just found it. I’ve already added my resource key as the API key, but it still doesn’t transcribe Portuguese. Would you be able to join a quick call? This is quite important for us and could be a key factor in selecting VAPI as our conversational AI provider.
b
Hey @Diogo Gonçalves what time zone are you residing in? I'll likely be available in a couple of hours or the next day. I am currently in the AEST time zone
d
Western European Time! And typically available between 8:30am and 6:30pm.
b
Gotcha, can you confirm if the language you need is on this list? https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=stt
pt-BR / pt-PT
d
it's pt-PT
I was already able to fine-tune a model and have a deployment ready
b
Got it, and BYOK did not work for you?
b
that's possible, we're adding in languages soon so that should solve the issue
h
Can you please add Czech (Czechia) - cs-CZ? Many thanks!
d
Do you have an ETA for adding those new languages? @bryant
b
@Diogo Gonçalves Should be out within the next week!
d
Perfect, keep me posted! 😆
b
Hey @Diogo Gonçalves , I believe it should be out now through the API
d
great, we will try it out!
m
hey @bryant I followed Diogo's instructions: inside Provider Keys > Transcriber > Azure Speech, I selected the region of my resource and entered my resource key under API Key. This didn't raise any errors. Then I started a call using the client SDK, setting the provider to azure and the language to pt-PT. The transcriber successfully transcribed what I said in Portuguese. However, when I inspected the logs of my endpoint, I noticed that Vapi made no requests to my endpoint. This means that BYOK is not working. AFAIK, Vapi needs to know the service region, the resource key, and the endpoint ID to connect to the fine-tuned model. Can you help me configure Vapi so I can use my fine-tuned STT model?
d
Hi @bryant! could you help here?
Hi @User @User could you help?
d
We would like to point to a fine-tuned model deployment and still leverage your integration. We were wondering whether it would be possible do it by setting the deployment key on provider keys (as @Miguel C detailed) but no requests are made to our endpoint. Any ideia if we can do it somehow? @Shubham Bajaj
s
@Diogo Gonçalves Yes, we do have the option to use Azure credentials for Azure provider. As you mentioned it's not working for you, could you share the recent call ID? https://cdn.discordapp.com/attachments/1320772372621430855/1327625930792894495/Screenshot_2025-01-11_at_18.40.04.png?ex=6783bf7f&is=67826dff&hm=098ba46f83bf01598e94fb3157b77b7477c0baa8028339c5a7057785c2d8f0e8&
d
Call ID: 46d9ed9b-aada-43d0-93a0-784e2e1d0ade
@Shubham Bajaj
s
@Diogo Gonçalves just checked your keys are used for this call.
d
What do you mean? I think we were already not able to convey our problem. We want to use a azure fine-tuned model - can we do it? we would need to pass a resource key, a region and a endpoint id. Providing an api key will never be enough