Speech service quotas and limits
This article contains a quick reference and a detailed description of the quotas and limits for the Speech service in Azure AI services. The information applies to all pricing tiers of the service. It also contains some best practices to avoid request throttling.
For the free (F0) pricing tier, see also the monthly allowances at the pricing page.
Quotas and limits reference
The following sections provide you with a quick guide to the quotas and limits that apply to the Speech service.
For information about adjustable quotas for Standard (S0) Speech resources, see more explanations, best practices, and adjustment instructions. The quotas and limits for Free (F0) Speech resources aren't adjustable.
Important
If you switch a Speech resource from Free (F0) to Standard (S0) pricing tier, the change of the corresponding quotas may take up to several hours.
Speech to text quotas and limits per resource
This section describes speech to text quotas and limits per Speech resource. Unless otherwise specified, the limits aren't adjustable.
Real-time speech to text and speech translation
You can use real-time speech to text with the Speech SDK or the Speech to text REST API for short audio.
Important
These limits apply to concurrent real-time speech to text requests and speech translation requests combined. For example, if you have 60 concurrent speech to text requests and 40 concurrent speech translation requests, you'll reach the limit of 100 concurrent requests.
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| Concurrent request limit - base model endpoint | 1 This limit isn't adjustable. |
100 (default value) The rate is adjustable for Standard (S0) resources. See more explanations, best practices, and adjustment instructions. |
| Concurrent request limit - custom endpoint | 1 This limit isn't adjustable. |
100 (default value) The rate is adjustable for Standard (S0) resources. See more explanations, best practices, and adjustment instructions. |
Batch transcription
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| Speech to text REST API limit | Not available for F0 | 100 requests per 10 seconds (600 requests per minute) |
| Max audio input file size | N/A | 1 GB |
| Max number of blobs per container | N/A | 10000 |
| Max number of files per transcription request (when you're using multiple content URLs as input). | N/A | 1000 |
| Max audio length for transcriptions with diarization enabled. | N/A | 240 minutes per file |
Model customization
The limits in this table apply per Speech resource when you create a custom speech model.
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| REST API limit | 100 requests per 10 seconds (600 requests per minute) | 100 requests per 10 seconds (600 requests per minute) |
| Max number of speech datasets | 2 | 500 |
| Max acoustic dataset file size for data import | 2 GB | 2 GB |
| Max language dataset file size for data import | 200 MB | 1.5 GB |
| Max pronunciation dataset file size for data import | 1 KB | 1 MB |
Max text size when you're using the text parameter in the Models_Create API request |
200 KB | 500 KB |
Text to speech quotas and limits per resource
This section describes text to speech quotas and limits per Speech resource.
Real-time text to speech
You can use real-time text to speech with the Speech SDK or the Text to speech REST API. Unless otherwise specified, the limits aren't adjustable.
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| Maximum number of transactions per time period for prebuilt neural voices. | 20 transactions per 60 seconds This limit isn't adjustable. |
200 transactions per second (TPS) (default value) The rate is adjustable up to 1000 TPS for Standard (S0) resources. See more explanations, best practices, and adjustment instructions. |
| Max audio length produced per request | 10 min | 10 min |
Max total number of distinct <voice> and <audio> tags in SSML |
50 | 50 |
| Max SSML message size per turn for websocket | 64 KB | 64 KB |
Audio Content Creation tool
| Quota | Free (F0) | Standard (S0) |
|---|---|---|
| File size (plain text in SSML)1 | 3,000 characters per file | 20,000 characters per file |
| File size (lexicon file)2 | 30KB per file | 100KB per file |
| Billable characters in SSML | 15,000 characters per file | 100,000 characters per file |
| Export to audio library | 1 concurrent task | N/A |
1 The limit only applies to plain text in SSML and doesn't include tags.
2 The characters of lexicon file aren't charged. Only the lexicon elements in SSML are counted as billable characters. Refer to billable characters to learn more.
Detailed description, quota adjustment, and best practices
Some of the Speech service quotas are adjustable. This section provides more explanations, best practices, and adjustment instructions.
The following quotas are adjustable for Standard (S0) resources. The Free (F0) request limits aren't adjustable.
- Speech to text concurrent request limit for base model endpoint and custom endpoint
- Text to speech maximum number of transactions per time period for prebuilt neural voices
- Speech translation concurrent request limit
Before requesting a quota increase (where applicable), check your current TPS (transactions per second) and ensure that it's necessary to increase the quota. Speech service uses autoscaling technologies to bring the required computational resources in on-demand mode. At the same time, Speech service tries to keep your costs low by not maintaining an excessive amount of hardware capacity.
Let's look at an example. Suppose that your application receives response code 429, which indicates that there are too many requests. Your application receives this response even though your workload is within the limits defined by the Quotas and limits reference. The most likely explanation is that Speech service is scaling up to your demand and didn't reach the required scale yet. Therefore the service doesn't immediately have enough resources to serve the request. In such cases, increasing the quota won’t help. In most cases, the Speech service will scale up soon, and the issue causing response code 429 will be resolved.
General best practices to mitigate throttling during autoscaling
To minimize issues related to throttling, it's a good idea to use the following techniques:
- Implement retry logic in your application.
- Avoid sharp changes in the workload. Increase the workload gradually. For example, let's say your application is using text to speech, and your current workload is 5 TPS. The next second, you increase the load to 20 TPS (that is, four times more). Speech service immediately starts scaling up to fulfill the new load, but is unable to scale as needed within one second. Some of the requests get response code 429 (too many requests).
- Test different load increase patterns. For more information, see the workload pattern example.
- Create more Speech service resources in different regions, and distribute the workload among them. (Creating multiple Speech service resources in the same region won't affect the performance, because all resources are served by the same backend cluster).
The next sections describe specific cases of adjusting quotas.
Speech to text: increase real-time speech to text concurrent request limit
By default, the number of concurrent real-time speech to text and speech translation requests combined is limited to 100 per resource in the base model, and 100 per custom endpoint in the custom model. For the standard pricing tier, you can increase this amount. Before submitting the request, ensure that you're familiar with the material discussed earlier in this article, such as the best practices to mitigate throttling.
Note
Concurrent request limits for base and custom models need to be adjusted separately. You can have a Speech service resource that's associated with many custom endpoints hosting many custom model deployments. As needed, the limit adjustments per custom endpoint must be requested separately.
Increasing the limit of concurrent requests doesn't directly affect your costs. The Speech service uses a payment model that requires that you pay only for what you use. The limit defines how high the service can scale before it starts throttle your requests.
You aren't able to see the existing value of the concurrent request limit parameter in the Azure portal, the command-line tools, or API requests. To verify the existing value, create an Azure support request.
Have the required information ready
- For the base model:
- Speech resource ID
- Region
- For the custom model:
- Region
- Custom endpoint ID
How to get information for the base model:
- Go to the Azure portal.
- Select the Speech service resource for which you would like to increase the concurrency request limit.
- From the Resource Management group, select Properties.
- Copy and save the values of the following fields:
- Resource ID
- Location (your endpoint region)
How to get information for the custom model:
- Go to the Speech Studio portal.
- Sign in if necessary, and go to Custom speech.
- Select your project, and go to Deployment.
- Select the required endpoint.
- Copy and save the values of the following fields:
- Service Region (your endpoint region)
- Endpoint ID
Create and submit a support request
Initiate the increase of the limit for concurrent requests for your resource, or if necessary check the current limit, by submitting a support request. Here's how:
- Ensure you have the required information listed in the previous section.
- Go to the Azure portal.
- Select the Speech service resource for which you would like to increase (or to check) the concurrency request limit.
- In the Support + troubleshooting group, select New support request. A new window appears, with auto-populated information about your Azure subscription and Azure resource.
- In Summary, describe what you want (for example, "Increase speech to text concurrency request limit").
- In Problem type, select Quota or Subscription issues.
- In Problem subtype, select either:
- Quota or concurrent requests increase for an increase request.
- Quota or usage validation to check the existing limit.
- Select Next: Solutions. Proceed further with the request creation.
- On the Details tab, in the Description field, enter the following:
- A note that the request is about the speech to text quota.
- Choose either the base or custom model.
- The Azure resource information you collected previously.
- Any other required information.
- On the Review + create tab, select Create.
- Note the support request number in Azure portal notifications. You're contacted shortly about your request.
Example of a workload pattern best practice
Here's a general example of a good approach to take. It's meant only as a template that you can adjust as necessary for your own use.
Suppose that a Speech service resource has the concurrent request limit set to 300. Start the workload from 20 concurrent connections, and increase the load by 20 concurrent connections every 90-120 seconds. Control the service responses, and implement the logic that falls back (reduces the load) if you get too many requests (response code 429). Then, retry the load increase in one minute, and if it still doesn't work, try again in two minutes. Use a pattern of 1-2-4-4 minutes for the intervals.
Generally, it's a good idea to test the workload and the workload patterns before going to production.