Whats good! I've got an pretty annoying problem with an AI for leads which stops talking when someone says a short sentence like maybe just "no" , or sometimes just doesnt answer, i dont really understand when it doesnt and why it doesnt some times, but usually when its a short sentence . A call id is: 16fd5bc2-c035-411f-aa05-46d512397332 ; as you can observe in the call, at 0:30 seconds it just doesnt say anything till i say something else. (The ai is in the language romanian)

Hey there! 👋 I can't access the specific details of your call at the 0:30 second mark, but I can definitely help explain why your Voice AI Caller might be stopping when it hears short sentences in Romanian! Based on the information I've found, there are several common reasons why this happens: ## Endpointing Issues The AI might be prematurely detecting the end of speech with short sentences like "no." According to [Discord discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1343636482874933340), the default Deepgram endpointing value of 10ms is often too aggressive, especially for short utterances. Increasing the

endpointing

value in the transcriber settings (e.g., to 300ms) and increasing

waitSeconds

in the

startSpeakingPlan

(e.g., to 0.8) could help. ## Transcriber Limitations with Romanian The AI may struggle specifically with Romanian language processing. As mentioned in [another thread](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1303981615344779325), "the transcriber might not be good enough for the language (Romanian)." Some users found success by switching transcribers - moving from Deepgram to another option. ## Configuration Settings Your AI's response timing can be adjusted through parameters like: - `numWordsToInterruptAssistant`: Setting this higher gives people more time to speak - `maxDurationSeconds`: Controls how long the AI waits for a response ## Other Potential Causes - The AI might be [hearing itself](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1337125226747662427) due to echoing - Short phrases might be misinterpreted as [end-call signals](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1337162299621900359) - There could be [model issues](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1339996587530915902) with OpenAI models not generating responses

how can i increase the endpointing value in the transcriber settings

@victor Here's the curl command to update the assistant configuration:

bash
curl --request PATCH \
  --url "https://api.vapi.ai/assistant/{assistant_id}" \
  --header "Authorization: Bearer {your_api_token}" \
  --header "Content-Type: application/json" \
  --data '{
    "transcriber": {
      "provider": "deepgram",
      "endpointLatencyMs": 300
    },
    "responseDelaySeconds": 0.6,
    "voice": {
      "provider": "livekit"
    }
  }'

## Values that need to be replaced: - `{assistant_id}`: Replace with the actual ID of the assistant you want to update - `{your_api_token}`: Replace with your Vapi API token This command will: 1. Set the Deepgram endpoint latency to 300 milliseconds 2. Set the response delay seconds (Start Speaking Plan latency) to 0.6 seconds 3. Change the voice provider from VAPI to LiveKit