I'm seeing the exact same issue, reproduced independently on two separate apps in my account, plus in the real product context that originally surfaced it.
Original context: I'm building a Quest 3 app in Unreal Engine using Voice SDK / wit-unreal. Testing in Unreal's VR Preview mode targeting Quest 3, I confirmed that voice input is captured correctly and the raw ASR transcription is returned as expected. I also confirmed (via the app's Understanding page, "Improve your app" audio review queue) that the recorded audio and its transcription ARE being sent to Wit.ai and show up as training data candidates. So the capture/transmission pipeline is working correctly end-to-end.
To rule out any client-side (UE/Quest) involvement, I isolated the issue entirely outside Unreal:
- App A: created 2026-08-31 (the original app used by the Unreal project)
- App B: created 2026-09-03 (fresh test app, created specifically to isolate this issue)
Steps taken on both apps:
- Added utterances (e.g. "おはよう" / "おはようございます") via the Understanding UI
- Assigned them to a simple intent (e.g. "greeting")
- Clicked "Train and Validate" on both the Understanding page and the Management > Utterances page — received the "utterance is now in the training queue" confirmation each time
- Waited 10+ minutes
- Queried GET /message directly via curl (no Unreal/Quest involved at all) with the exact trained utterance text, tried both v=20260903 and v=20240304
Result: intents is always [], entities is always {}, despite text matching the trained utterance exactly. Both apps otherwise return valid HTTP 200 responses with the correct text field, so auth and basic request handling work fine — only the NLU inference itself appears to return nothing.
This strongly suggests a backend issue affecting newly created apps (both of mine were created within the last ~4 days of testing), consistent with what's described in this issue and also #2825. Happy to share app IDs privately if that helps investigation.
I'm seeing the exact same issue, reproduced independently on two separate apps in my account, plus in the real product context that originally surfaced it.
Original context: I'm building a Quest 3 app in Unreal Engine using Voice SDK / wit-unreal. Testing in Unreal's VR Preview mode targeting Quest 3, I confirmed that voice input is captured correctly and the raw ASR transcription is returned as expected. I also confirmed (via the app's Understanding page, "Improve your app" audio review queue) that the recorded audio and its transcription ARE being sent to Wit.ai and show up as training data candidates. So the capture/transmission pipeline is working correctly end-to-end.
To rule out any client-side (UE/Quest) involvement, I isolated the issue entirely outside Unreal:
Steps taken on both apps:
Result:
intentsis always[],entitiesis always{}, despitetextmatching the trained utterance exactly. Both apps otherwise return valid HTTP 200 responses with the correcttextfield, so auth and basic request handling work fine — only the NLU inference itself appears to return nothing.This strongly suggests a backend issue affecting newly created apps (both of mine were created within the last ~4 days of testing), consistent with what's described in this issue and also #2825. Happy to share app IDs privately if that helps investigation.