Video Translation REST API: first iteration intermittently waits ~6 minutes in NotStarted, while Speech Studio on the same resource starts immediately

The six-minute NotStarted interval is not documented as expected behavior for Video Translation API 2026-03-01, so preserve the affected operation details and open an Azure support request for service-side investigation rather than changing the App Service configuration.

What the status indicates

For Video Translation, an iteration normally progresses from NotStarted to Running, then to Succeeded or Failed; the documented client workflow is to poll until a terminal state. The iteration response also exposes createdDateTime and lastActionDateTime, which can be used to measure how long it remains in each state.^1^^2^

Microsoft's documentation does not currently define a six-minute startup window, inactivity timeout, cold-start interval, or latency guarantee for this API. The published Speech known-issues list also does not identify this Video Translation symptom.^3^

Your measurements are consistent with backend queueing or capacity activation, but they do not establish the exact cause. Azure Speech generally uses on-demand autoscaling and can experience latency while capacity scales, although the quota documentation does not specifically confirm that mechanism for Video Translation.^4^

Because both the metadata and fully automatic paths exhibit the same delay, changing firstIterationInput or MetadataJson is unlikely to address the admission delay.

What can reduce the delay

There is no documented Video Translation setting for:

• disabling idle scale-down;
• keeping the backend warm;
• reserving or purchasing dedicated Video Translation capacity;
• increasing a Video Translation queue quota;
• obtaining guaranteed start latency.

The adjustable Speech quotas listed by Microsoft cover specific capabilities such as speech-to-text, fast/batch transcription, speech translation, and text-to-speech; Video Translation is not listed as an adjustable quota. Microsoft also notes that increasing a quota does not necessarily resolve autoscaling delays.^4^

Always On only keeps your Linux App Service worker active. Since your worker submits immediately and the iteration is successfully created as NotStarted, changing or scaling the B1 App Service would not remove queueing after acceptance by Video Translation.

West Europe is a supported Video Translation region.^5^ Testing a second supported region could identify whether the behavior is regional, but it is not a documented fix. Use a separate Speech resource and identical input, then compare several first-after-idle and immediate-follow-up jobs. Consider data residency and duplicated-resource configuration before changing production traffic.

Avoid periodic dummy translations as a production workaround unless Microsoft Support explicitly recommends them: no supported keep-warm operation is documented, and real translation requests can incur usage charges.

Support request evidence

Open Azure portal → Help + support → Create a support request against the Speech resource. Include:

1. Resource region and pricing tier.
2. API version 2026-03-01.
3. Translation ID and iteration ID for each example.
4. The unique Operation-Id used for every create operation.
5. Submission time in UTC, createdDateTime, and every observed lastActionDateTime.
6. Exact NotStarted → Running → terminal durations.
7. Sanitized request and response bodies, plus response headers/correlation IDs.
8. A paired example showing approximately 360–373 seconds for the first job and 3–10 seconds for the immediately following job.
9. The inactivity duration preceding each slow job.
10. Confirmation that the same pattern occurs with and without supplied metadata.

This lets the service team distinguish regional capacity scheduling from a product defect and confirm whether another capacity arrangement is available outside the publicly documented options.

Billing

The public pricing page meters Video Translation by input-video hours and output-video hours, with separate output rates for standard and personal voice; it does not describe billing by wall-clock queue duration.^6^

Therefore, the six minutes in NotStarted should not itself become six minutes of additional media usage under the published meters. However, the documentation does not explicitly state the billing treatment of NotStarted. Verify the actual charge in Cost Management + Billing → Cost analysis, grouped by meter, and ask Support to confirm the meter records for the affected operation IDs if the charged quantity exceeds the input/output media duration.

References

1. How to use video translation - Speech service - Foundry Tools
2. Translation Operations - Get Translation - REST API (Azure Azure AI ...
3. Azure Speech in Foundry Tools known issues - Foundry Tools
4. Quotas and limits for Azure Speech
5. Supported regions for Azure Speech
6. Azure Speech in Foundry Tools pricing