Issues with the virtual assistant model using a custom keyword

Question

I've been trying to create a voice assistant model with a custom keyword since yesterday, but it's still stuck in “processing” status. I set up the basic model, which they say takes a few hours at most, if not 30 minutes at the latest. How is it possible that after 24 hours it’s still processing? Another model, on the other hand, has been processing for 12 hours

Answer 1

Hello Damiano Buonaugurio,

Welcome to Microsoft Q&A .Thank you for reaching out to us.

The observed behavior indicates that custom keyword (wake word) model builds are remaining in a “Processing” state for extended durations (12+ to 24+ hours) across multiple attempts.

Under normal conditions, keyword model creation—especially for basic models—usually completes within minutes after submission. However, model creation is handled asynchronously, and there are cases where processing takes significantly longer due to backend conditions. When multiple models remain in progress simultaneously without failure or error, it is typically an indication that the requests have been accepted but are delayed within the processing pipeline.

The following factors most commonly contribute to this situation -

1. Regional capacity constraints - High demand in a selected region can delay the execution of model creation jobs.
2. Backend queue delays - Model compilation requests are queued and processed in sequence, which may result in extended wait times during periods of heavy load
3. Transient service-side conditions - When more than one model is affected, the behavior is typically aligned with a temporary service-side delay rather than an issue in the setup itself

Please check if the following suggestions help -

1. Validating model setup
   1. Please confirm the correct model type (Basic or Advanced)
   2. Basic models are expected to complete faster compared to advanced configurations
2. Retrying in a different region
   1. Please create a Speech resource in another supported region
   2. Then recreate the keyword model in that region
   3. Region-based delays are a common contributing factor
3. Please avoid concurrent submissions
4. Limit model creation to one request at a time
   1. Parallel submissions can increase queue delays
5. Recreate the model when delay is prolonged - If the model remains in processing for an extended duration (for example, ~24 hours), creating a new model is generally more effective than waiting further
6. Refreshing and validating model status
   1. Please reopen the model page in Speech Studio
   2. Then confirm that the displayed status reflects the current backend state
7. Checking service health signals - Reviewing the Azure Service Health section in the portal to identify any active service conditions affecting processing

The following references might be helpful , please check them out

• Keyword recognition overview - Speech service - Foundry Tools | Microsoft Learn
• Create a custom keyword quickstart - Speech service - Foundry Tools | Microsoft Learn

 

Thank you

Answer 2

@Karnam Venkata Rajeswari
Thank you so much for your reply.
I did notice that using other regions generated the model more quickly. Now that I’ve checked, I see that the first models have also been completed. Instead of taking a few minutes or a few hours, they took a few days.