We're running an outbound calling campaign and experiencing consistent voicemail detection failures with a specific pattern. The problem: Carrier-based voicemail systems that read the phone number aloud (e.g., "4 5 6 5 3 1 8 8 7 is not available") are not being detected as voicemail. The transcriber picks this up as user speech, and our assistant starts conversing with the voicemail system. Example call ID: 019b2e8b-31a2-755f-91f0-9812c0c41719 What we've tried: All voicemail detection providers: Vapi, Twilio, Google, OpenAI — none catch this Twilio voicemailDetectionTypes settings — caused worse issues (random hangups) Voicemail tool with prompt instructions to detect number sequences — LLM can't process fast enough before responding Structured system prompts with explicit voicemail indicators Current wait Seconds is 2.5s, but increasing further creates awkward pauses for real humans Current config: Assistant ID: 7b9c01f4-2870-42c8-a161-00ecdc293461 Voicemail detection: provider: "vapi" voicemailMessage is configured and works when detection triggers The ask: Is there a way to improve detection for this specific pattern (phone numbers being read aloud by carrier voicemail)? Or is this a known gap in the detection models?

Try switching to the Google provider with these settings for improved detection of carrier voicemail systems that read phone numbers aloud:

json
{
  "voicemailDetection": {
    "provider": "google",
    "backoffPlan": {
      "startAtSeconds": 1.5,
      "frequencySeconds": 2.5,
      "maxRetries": 10
    }
  }
}

This configuration increases detection persistence and speed, which may help catch those edge-case voicemail greetings. For more details, see the [Vapi voicemail detection documentation](https://docs.vapi.ai/calls/voicemail-detection). Source: - [Vapi Voicemail Detection Documentation](https://docs.vapi.ai/calls/voicemail-detection)

Should we keep the voicemail tool as well? I know we tried google's detection before and we missed normal voicemails where people recorded their own voices.

I use Vapi's Detection + Provide a voicemail_flag tool to the assistants. Important to add to the prompt that it should utilize the tool whenver it recognizes voicemail.

I think the issue is that even with thorough system prompts that give instructions to flag things as voicemails, in those first 2-3 seconds, it doesnt recognize the recording as VM fast enough and interrupts the VM recording. We have a VM tool as well, but in the cases where the recording doesnt follow "traditional VM format", the assistant starts speaking after a few seconds.