TTS(Elevenlabs) Buffer increase? Vapi AI #support

TTS(Elevenlabs) Buffer increase?

Ahmed

04/12/2026, 1:46 PM

is there any way to increase the number of characters/tokens buffered before sending to Elevenlabs? I'm running my LLM on openrouter and sometimes the tokens stream slowly which makes the TTS sound a bit disjointed. I'm using Arabic, already maxed input min characters at 80, and disabled Auto mode if that helps. I'm using Flash v2.5, and it sounds ok when used standalone (from elevenlabs' website) but falls apart in call. Example call: 019d81bc-82ca-7ee4-b41b-dd240c3f9255

Chiranjeet Mishra

04/12/2026, 1:46 PM

Hi there, Thank you for your message. Our team is currently out of the office. We operate Monday through Friday, from 9:00 AM to 8:00 PM Pacific Standard Time (PST). We’ll get back to you as soon as possible during our normal business hours. If your message is urgent, please mark it accordingly or include “URGENT” in the subject line, and we’ll do our best to respond promptly. Warm regards, Vapi Customer Support Team

Chiranjeet Mishra

04/18/2026, 3:34 AM

Hi there, I reviewed the call log for

019d81bc-82ca-7ee4-b41b-dd240c3f9255

and wanted to share a few findings. **The character buffering (minCharacters: 80
) isn’t the issue.** Your LLM is generating tokens very quickly, so responses are completing before the threshold is hit. As a result, everything is being sent to ElevenLabs as a single block—no mid-stream chunking is actually happening. The main issue is the 0ms endpointing timeout, which is causing duplicate model requests. The transcriber sends a partial transcript, immediately triggering an LLM request, then the final transcript arrives milliseconds later and triggers another. The first response gets discarded, adding ~500–1000ms delay per turn. Recommended changes: 1. Increase endpointing timeouts (highest impact):

Copy code

"startSpeakingPlan": {
  "waitSeconds": 0.7,
  "transcriptionEndpointingPlan": {
    "onPunctuationSeconds": 0.1,
    "onNoPunctuationSeconds": 0.8,
    "onNumberSeconds": 0.5
  }
}

Even a small delay (e.g. 100ms on punctuation) prevents most duplicate requests, and

onNoPunctuationSeconds: 0.8

avoids triggering on partial transcripts. 1. Lower

minCharacters

for Arabic (minor improvement):

Copy code

"voice": {
  "chunkPlan": {
    "enabled": true,
    "minCharacters": 30
  }
}

Arabic responses tend to be shorter, so this is a better fit. It won’t significantly impact latency with your current setup, but helps in cases with longer outputs. Lastly, the transcript fragmentation (multiple

AI:

lines) is just from ElevenLabs sync alignment. It doesn’t reflect actual audio gaps. Best, VAPI Support

Previous Next