Azure Speech fast transcription returns HTTP 429 despite very low usage

Kiran Roberts 0 Reputation points
2026-11-18T19:02:09.6566667+00:00

I'm developing a controlled prototype using Azure Speech fast transcription.

Resource configuration

Pricing tier: S0 Standard

Region: UK South

API: Azure Speech fast transcription

Recognition locale: Albanian (sq-AL)

Problem

Two individually authorised transcription requests returned HTTP 429, on 8 November and 12 November 2026.

Azure Metrics recorded one Total Call and one Blocked Call in the relevant monitoring interval on 12 November.

An earlier English transcription request succeeded using the same application workflow.

Could this indicate regional fast-transcription capacity or resource-specific throttling? Are there recommended checks to identify why these isolated requests were rejected?

No automatic retries were performed, and the application has disabled further attempts on the affected recording.

Additional context from an earlier report

I previously reported similar HTTP 429 failures on 19 and 26 October 2026, using the same Azure Speech service.

A previous Microsoft Q&A AI-generated answer suggested exponential retries and possible idle-to-active scaling. However, the underlying throttling or capacity cause remains unconfirmed.

Our application handles public-sector communication and requires deliberate authorisation before external audio processing. We therefore cannot introduce unrestricted automatic retries or warm-up requests.

Could Microsoft clarify:

Whether Albanian fast transcription (sq-AL) has any relevant backend capacity constraints in UK South.

How to distinguish resource-level throttling from temporary service-capacity limitations.

Which non-sensitive request identifiers, response headers or diagnostic logs would help establish the precise cause without sending another transcription request.

Azure API Management
Azure API Management

An Azure service that provides a hybrid, multi-cloud management platform for APIs.

0 comments No comments

1 answer

Sort by: Most helpful
  1. AI answer

    2026-11-18T19:03:51.0466667+00:00

    The existing evidence cannot distinguish resource throttling from temporary backend capacity, so preserve the failed-response metadata, correlate Azure Monitor and Service Health for the exact UTC windows, and open a Speech resource support request without resubmitting the recording.

    What the current evidence establishes

    • Low usage does not exclude a service-capacity 429. Fast transcription and batch transcription share an S0 limit of 600 requests per minute, but Microsoft documents that Speech can return 429 while a workload remains within quota when autoscaling has not yet supplied enough compute. Increasing quota does not resolve that condition.^1^
    • One Total Call plus one Blocked Call identifies the rejected request, not its underlying cause. BlockedCalls means calls that exceeded a rate or quota limit. Its available dimensions are ApiName, OperationName, Region, and RatelimitKey.^2^
    • UK South supports fast transcription. The regional availability table explicitly marks uksouth as supported.^3^
    • The published material does not identify an Albanian-specific fast-transcription capacity restriction in UK South. Therefore, the repeated sq-AL pattern is useful evidence for Microsoft Support, but it does not by itself prove locale-specific saturation.
    • There is a prior public report matching S0, UK South, sq-AL, and isolated 429 responses, but no corresponding publicly confirmed regional incident was identified.^4^

    Checks that require no new transcription

    1. Break down the existing metrics
      • Azure portal → Speech resource → Monitoring → Metrics.
      • Set each affected UTC interval narrowly around the request.
      • Add Blocked Calls, Total Calls, Client Errors, and Server Errors where available.
      • For Blocked Calls, use Apply splitting for:
        • ApiName
        • OperationName
        • Region
        • RatelimitKey
      • Export or capture the chart and dimension values. Metric filtering and splitting are intended to identify which segment contributed to a result.^5^
      If the relevant minute contains one call and one blocked call, with no surrounding traffic, that argues against reaching the documented 600-RPM limit through this application. It still cannot separate an internal rate-limit key from autoscaling or backend-capacity throttling.
    2. Review subscription-specific health history
      • Azure portal → Service Health → Service issues.
      • Filter for Azure AI/Speech, UK South, and the affected subscription.
      • Check both active and resolved events covering:
        • 19 and 26 October 2026
        • 8 and 12 November 2026
      • Record any tracking ID, UTC start/end time, impacted region/resource, and issue updates.
      Service Health is subscription- and region-specific, unlike the public Azure Status page, which focuses on broad incidents.^6^
    3. Check Resource Health
      • Speech resource → Resource health → Health history.
      • Review the same UTC windows.
      Resource Health reports current and historical service problems affecting an individual Azure resource.^7^
    4. Review previously enabled diagnostic logs
      • If a diagnostic setting was already active, query or export the records for the exact timestamps.
      • Microsoft.CognitiveServices/accounts supports RequestResponse, Audit, and Trace log categories.^8^
      • If diagnostics were not enabled at the time, enabling them now will help future failures but will not reconstruct old request records.
    5. Preserve client-side evidence already captured
      • Exact UTC timestamp, including milliseconds and timezone.
      • HTTP status, complete response body, and all response-header names and values except secrets.
      • Resource name and resource ID, region, pricing tier, operation, API version, locale, audio size/duration/format, and authentication method.
      • Application-generated trace or correlation ID, if one was attached.
      • Evidence that no concurrent fast or batch transcription calls occurred.
      The available Fast Transcription documentation does not specify a guaranteed diagnostic request-header name for this 429 response, so do not invent or rely on a particular correlation header. Supply every non-sensitive identifier actually returned.

    Escalation

    From the Speech resource, select Support + Troubleshooting and create a technical support request. Include the four failure timestamps, metric exports with dimensions, full sanitized 429 responses, Service/Resource Health findings, and the successful English request as a control case. Azure Support can perform service-side correlation that customer-visible metrics cannot.^9^

    Ask Support to determine explicitly whether each request was rejected by:

    • the resource’s shared fast/batch request-rate quota;
    • an internal RatelimitKey;
    • Speech autoscaling or regional backend capacity;
    • model/locale-specific capacity for sq-AL in UK South.

    No warm-up call or unrestricted retry is required for this investigation. For future authorised attempts, Microsoft recommends bounded retries for 429, but that is an operational mitigation rather than proof of the cause.^10^


    References

    1. Quotas and Limits for Azure Speech - Foundry Tools | Microsoft Learn
    2. Supported metrics for Microsoft.CognitiveServices/accounts
    3. Supported regions for Azure Speech
    4. Azure Speech Fast Transcription repeatedly returns HTTP 429 despite very low usage — UK South, S0 - Microsoft Q&A
    5. Analyze metrics with Azure Monitor metrics explorer
    6. Azure Service Health FAQ
    7. Resource Health overview
    8. Supported logs for Microsoft.CognitiveServices/accounts
    9. Create an Azure support request
    10. Use the fast transcription API - Speech service - Foundry Tools | Microsoft Learn
    AI-generated content may be incorrect. Read our transparency notes for more information.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.