Missing audio files when creating synthetic audio for testing transcripts in Speech Studio

Question

In order to test a custom language model (with a paid subscription), we tried creating synthetic audio for 37 test sentences for a test dataset.
Even though the report showed 0 errors, only 31 audio files had been generated.
Trying to find the reason why some lines would not be processed, I sorted the test sentences and re-tried creating the test dataset.
This time, only 25 audio files were created, even though using the exact same set of sentences, just in a different order!
Once again, the report showed 0 errors (but only 25 times success).

Looking at the JSON file gave a bit more information in that there were several errors during the creation of audio files (but those errors didn't appear in the report!):
Most of the time the error was "Connection was closed by the remote host. Error code: 1013. Error details: Status(StatusCode=\"ResourceExhausted\", Detail=\"No free synthesizer\") USP state: 3. Received audio size: 0bytes." and one time just "USP error, timeout to get the first audio chunk."

The region is France Central, pricing tier S0.

I'm completely lost on why creation of those audio files would fail silently and what to do about it (apart from re-uploading the same file again and again, hoping for as many files as possible to be created and blowing a lot of resources in the process).

Please advise.

Answer 1

Your SSML contains error. I had this issue before. Debug the SSML and avoid bad limitations