Hi Vapi team, We’re testing a Spanish phone assistant. We switched to Deepgram Nova 3 (Spanish) as suggested, but we still see intermittent cases where the user speaks (clear audio, even repeated “hola hola / sí”) and the assistant does not react — it looks like user speech is not being detected/forwarded to the model (ASR/VAD/endpointing issue). Symptoms: 1) User says a short/low-volume “sí” (audio clearly present), but transcript is missing and the assistant stays silent indefinitely. 2) Repeated greeting sometimes occurs (“How do you want to start?” asked twice). We suspect this may be related to tool-based “aperturas” snippets plus assistant’s own opening. 3) Tool calls (ragSearch) are logged sometimes but not consistently for similar user intents. We are trying to enforce 1 tool call at specific checkpoints. What we have already done: - Switched transcriber from OpenAI to Deepgram Nova 3 Spanish. - Observed minor packet loss in one call, but audio still sounds clear. - Added topic enums (amor/trabajo/familia/espiritual/general/precio/tiempo/retencion) and validated tool endpoint works. Questions / requests: A) For missing short utterances (“sí” low volume), what are your recommended settings for Deepgram confidence filtering and VAD/endpointing? - Is there a known best practice for confidenceThreshold (we currently used 0.4 before; we can lower to 0.1/0.0 for tests)? - Which call config fields should we adjust (vadAssistedEndpointingEnabled, endpointing thresholds, interruption/barge-in, etc.)? B) Can you confirm whether Talk-to-Assistant runs the exact same ASR/VAD pipeline as real phone calls? If not, what are the differences? C) For repeated greeting, do you have any guidance to avoid “double opening” when the assistant uses an “aperturas” snippet plus its own greeting? Call IDs to inspect: - (missing “sí” / silence) - (repeated greeting) - (control call) Thanks!

To improve detection of short/low-volume utterances like “sí” in Spanish, try this configuration for your assistant’s voice pipeline:

json
{
  "startSpeakingPlan": {
    "transcriptionEndpointingPlan": {
      "onPunctuationSeconds": 0.1,
      "onNoPunctuationSeconds": 1.5,
      "onNumberSeconds": 0.5
    },
    "waitSeconds": 0.4
  },
  "stopSpeakingPlan": {
    "numWords": 0,
    "voiceSeconds": 0.2,
    "backoffSeconds": 1.0
  },
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-3",
    "language": "es",
    "confidenceThreshold": 0.1
  }
}

Set

numWords: 0

and lower

confidenceThreshold

to 0.1 for more sensitive VAD-based interruption, which helps catch brief responses like “sí” ([see documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration)). Source: - [Voice pipeline configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration)

For short/low-volume utterances, we usually lower confidenceThreshold to 0.1–0.2 and enable vadAssistedEndpointing; also check interruptionEnabled for barge-in support. Talk-to-Assistant uses the same ASR/VAD pipeline as real calls, but live phone conditions (network jitter, packet loss) can still affect detection. For repeated greetings, consider disabling one of the “aperturas” or your assistant’s opening in those flows. Can you share a sample snippet of your greeting logic so we can pinpoint the double-trigger issue? @Estefan

Hi Estefan, Could you please share the call IDs where you’re seeing these issues? That will allow us to pull the logs and investigate the ASR/VAD behavior in detail. Once we have those, we can take a closer look and follow up with next steps. Best, VAPI Support