What was going on
Since October 12 our production text to speech broke. We convert SSML with the Speech SDK and now get SynthesizeCanceled with ServiceTimeout and USP error: timeout waiting for the first audio chunk, plus WS_OPEN_ERROR_UNDERLYING_IO_OPEN_FAILED. Nothing changed on our side.
What fixed it
We confirmed this is caused by the October 2022 Windows security updates, which changed TLS behavior and broke the SDK connection to the TTS service. The fix is to install the out of band update for your build: KB5020435 for Windows 10 20H2 through 22H2, KB5020387 for Windows 11 21H2, and KB5018496 for Windows 11 22H2. These rolled into the November automatic update cycle, so if you are on an unpatched machine the simplest fix is to check for updates. One note: speech recognition kept working through all this, only synthesis was affected. Source: https://github.com/Azure-Samples/cognitive-services-speech-sdk/issues/1692