Hi Vapi Support, We are testing a Spanish voice assistant (inbound DID via Netelip.com). We’re seeing three recurring issues: 1) Voice prosody instability (ElevenLabs Turbo v2.5): - sudden pitch spikes, sudden slowdowns (“stretched” speech), and sometimes syllable-by-syllable delivery (choppy / segmented). - This happens intermittently and breaks naturalness. Current voice settings (Vapi voice tab): - Stability: 0.6 - Similarity/Clarity: 0.5 - Speed: 0.9 - Optimize streaming latency: 2 - Input min characters: 15 - Background sound: off 2) Missed short utterances leading to “assistant goes silent”: - Example: user says a low-volume “sí” (yes). The transcript sometimes does not finalize, assistant waits silently until user repeats “hola”. We switched to Deepgram Nova 3 (Spanish), background denoising enabled, confidence threshold 0.1. We still see this issue; in some calls there is minor packet loss reported. 3) Call not ending when the caller hangs up: - User hangs up, but Vapi call remains active and only ends via “silence timeout” ~10 minutes later. This suggests Vapi may not be receiving BYE / disconnect from the SIP side. Please advise: - Recommended endpointing/VAD/turn-taking settings for Spanish to reliably detect short answers (“sí/no”). - Any recommended Deepgram parameters or alternatives (Nova 2 vs Nova 3) for better finalization. - Best practice voice streaming settings to avoid choppy syllabic prosody with ElevenLabs. - How to debug/resolve the missing hangup (BYE) / call disconnect issue with inbound SIP. Call examples (please check logs): - Call ID: [CALL_ID_1] (missed “sí”, assistant silent) - Call ID: [CALL_ID_2] (voice prosody spikes) - Call ID: [CALL_ID_3] (hangup not detected; ended by silence timeout) Thanks!

To improve detection of short Spanish utterances like "sí" or "no", try the following voice pipeline configuration for your assistant:

json
{
  "startSpeakingPlan": {
    "transcriptionEndpointingPlan": {
      "onPunctuationSeconds": 0.1,
      "onNoPunctuationSeconds": 1.5,
      "onNumberSeconds": 0.5
    },
    "waitSeconds": 0.4
  },
  "stopSpeakingPlan": {
    "numWords": 0,
    "voiceSeconds": 0.2,
    "backoffSeconds": 1.0
  }
}

This setup uses fast VAD-based interruption and text-based endpointing, which is recommended for reliably capturing short answers in Spanish conversations. Adjust

voiceSeconds

if you need even faster interruption or to reduce false positives. For more details, see the [voice pipeline configuration documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration). Source: - [Voice pipeline configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration)

Thanks for the detailed report, these pain points make sense and I can help troubleshoot. For prosody spikes, we can review ElevenLabs streaming and stability settings; for short utterances, tuning VAD and Deepgram confidence or testing Nova 2 might help. Regarding hangup detection, we can check SIP BYE signaling and endpoint settings. Can you share a snippet of your SIP logs around disconnect so we can pinpoint the missed hangups? @Estefan

Hi — thanks. On prosody spikes: we’re using ElevenLabs voice “[Marcela-2]” via Vapi streaming. Current settings in Vapi/Eleven Labs are (first vapi / 2nd )= 11labs): - Model: Eleven Turbo v2.5 / Eleven Turbo v2.5 - Stability: 0.6 / 65% - Similarity: 0.5 / 71% - Style / Speaker Boost / Auto mode: OFF - Speed: 0.9 / 1.00 - Style Exaggeration: 0 / None - Optimize Streaming Latency: 2 / None (tried many possibilities) - Confidence Threshold: 0.1 / None - Use Numerals: Yes / NOne - Background Denoising Enabled: Yes (tried as well: NO) 1. We see intermittent pitch/speed jumps and occasional “monosyllabic” delivery (MO-NO-SI-LA-BI-C). 2. Do you have a recommended ElevenLabs + Vapi streaming preset for telephony to reduce prosody instability (buffering/streaming knobs on Vapi side, or recommended stability/speed ranges)? 3. On missed short utterances (“sí” spoken softly): Deepgram Nova 3 General (Spanish), confidence threshold currently [0.1]. The transcript sometimes shows the short utterance but the agent doesn’t respond, or the call stalls after a user response. We want to run an A/B test: - but no way to swap Deepgram Nova 3 → Nova 2 (problem with select keyterms..). So no way to test 5 short calls, same script (soft “sí”, “izquierda/centro/derecha”, 3 numbers). Questions: 1) Which VAD/endpointing parameters in Vapi should we tune specifically to capture very short/low-volume utterances in telephony? 2) Do you recommend Nova 2 for this case, and if so, any specific config? Please, Help - Urgent Re hangup detection: Seems to be fixed...

From what you’re describing, the prosody issues and the missed “sí” are most likely coming from a mix of very aggressive low-latency streaming and conservative turn detection, especially over telephony where jitter and packet loss are common. When those combine, you get unstable speech and cases where the system hesitates to finalize short answers. My usual approach here is to first make sure the ElevenLabs output is fully stable, then fine-tune VAD and endpointing so soft, short replies are picked up quickly, and finally validate Nova 2 vs Nova 3 with a few controlled test calls. That way we’re not guessing or randomly changing settings, but actually isolating what’s causing the behavior. Some of this depends a lot on your exact Vapi and SIP setup, so it’s much easier to walk through it together. Let’s continue this conversation more privately and I’ll help you get it dialed in properly. @Estefan