Hey everyone! Building a multilingual Squad setup and running into a frustrating edge case with Gladia and VAD endpointing. Assistant 1 asks for a language preference. If the user gives a very short, single-word answer like "Türkçe" or "English", the utterance is completely swallowed/dropped. It seems the VAD filters it out as noise (since it's <500ms) before it even reaches the transcriber. What we tried (and the API/UI sync issue): To force the STT to catch these specific words, we tried injecting a Custom Vocabulary for Gladia via the API (since the Vapi UI doesn't expose vocabulary fields for Gladia yet). We successfully patched the assistant using this payload: { "transcriber": { "provider": "gladia", "model": "solaria-1", "customVocabularyEnabled": true, "customVocabularyConfig": { "vocabulary": ["Türkçe", "English"] } } } The API accepts this (200 OK). However, we are facing two blockers: UI Overwrite: The Vapi Dashboard flags this as "Unsaved changes." If we modify any other prompt/setting and hit Save on the UI, it strips our custom API config entirely. VAD Still Dropping: Even when the API config is actively running in the background, the VAD still swallows those short 1-word utterances. My Questions for the team/community: Is there an officially supported way to persist Gladia custom vocabularies without the Dashboard UI overriding it? What are the recommended VAD or silenceTimeoutSeconds tweaks to ensure short 1-word answers aren't dropped as background noise? (Trying to avoid using DTMF/Keypad to keep the experience purely voice-native). Thanks in advance for any insights!

To help prevent VAD from dropping short, single-word utterances, try lowering the

voiceSeconds

threshold in your

stopSpeakingPlan

for higher sensitivity. For example:

json
"stopSpeakingPlan": {
  "numWords": 0,
  "voiceSeconds": 0.1,  // More sensitive to short utterances
  "backoffSeconds": 1.0
}

Test this setting to balance responsiveness and background noise sensitivity. (See the voice pipeline configuration documentation for details.) Source: - [Voice pipeline configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration)

Yeah this is frustrating, but the issue isn’t Gladia vocab the audio is getting dropped at VAD before it even reaches STT. Right now the Vapi UI will overwrite anything it doesn’t support, so API configs like custom vocab won’t stick unless you avoid editing in the dashboard. The real fix is loosening VAD (lower endpointing sensitivity, bump silence timeout) so those short words actually pass through. I can help you tune this properly, are you running pure Vapi VAD or mixing it with Gladia’s endpointing? @Deadroom

Message marked as helpful by @Deadroom! 🎉