Unified ticket - Context We want to unify these 3 threads into a single investigation because symptoms overlap (VAD/endpointing thresholds, telephony conditions, and turn detection). Please close/merge the other tickets and continue here. Environment We want to focus this ticket on 2 issues in real phone calls (Netelip BYO SIP trunk, DID +34 910 05 30 14 , E164 ON). Voices: ElevenLabs (Marcela-2, Samuel). Planned ASR test: Deepgram Nova 2 Phonecall vs Nova 3. ==> no way to change from Nova 3 Genral to Nova 2 (error messages in Key.terms). 1) Prosody instability (ElevenLabs) * Marcela sometimes sounds slowed down and has sudden intonation “jumps” (Samuel also, less noticeable). We tried: * Marcela-2: stability 0.6→0.7; speed 0.9→0.8 * Samuel: stability 0.6→0.7; speed 1.1→1.0 * Is there any Vapi-side low-latency streaming/chunking/buffering setting that can cause this wobble? Recommended stable baseline? 2) Missed short/soft utterances + silence after loud noise We miss low-volume monosyllables (“sí/no”) and, if there’s a loud background noise spike (e.g. TV), the assistant can become unresponsive / no user turn created. Users won’t be in quiet rooms. Need from Vapi A) Exact recommended telephony settings to catch “sí/no”: confidenceThreshold (0.1–0.2?), vadAssistedEndpointing, endpointing timings (end-of-speech delay etc.), and interruptionEnabled. B) Mitigations for loud-noise spikes so speech still confirms (AGC/noise suppression/VAD mode guidance).

To improve detection of short utterances like "sí/no" in telephony environments, use VAD-assisted endpointing with a low confidence threshold and fast interruption settings. Here’s a recommended configuration snippet:

json
{
  "transcriber": {
    "confidenceThreshold": 0.1,
    "vadAssistedEndpointingEnabled": true
  },
  "stopSpeakingPlan": {
    "numWords": 0,
    "voiceSeconds": 0.2
  }
}

This setup ensures the assistant quickly reacts to brief, low-volume responses in noisy conditions. For more details, see the voice pipeline configuration documentation. Source: - [Voice pipeline configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration)

Please need help. This matter makes that calls are not "human like" ...

Sorry for the trouble here. Could you please provide the call IDs where this issue occurred? Once we have those, we can review the logs and investigate further on our end. Thanks, VAPI Support