Hello, Lately, the transcriber nova-2-phonecall (Deepgram) hasn’t been performing well. Here are some examples of incorrect transcription: - Difficulty recognizing the number of people for a booking: What was actually said: "2 person, 1,2 2 person, 2 people." What the transcriber captured: "2001, 2, 2002 people." VAPI Call ID: 525fb616-ae22-4412-b390-34b310972ea8 Time Code: 2:11 - Critical error—missed a keyword essential for the agent's script: Expected word: "existing" VAPI Call ID: 38d63916-3937-4f61-9592-4315811bc1fe Time Code: 0:20 - Entire conversation problematic—misheard key details: Transcriber misheard "birthday" as "Thursday" and failed to recognize the name "Madoka Yoshikawa" despite it being spoken clearly. VAPI Call ID: fb8c73f8-4da0-41d9-a70b-87439f2faf17 Time Code: 0:28 All of these issues occurred when there were no reported incidents on Deepgram's side. I can compile more evidence if needed to show that transcription issues are frequent and critical. So here’s my question—could you recommend an alternative transcriber model or provider? We’d really appreciate your advice! 😊

HeyKarina Krinitskaya Thank you for creating the ticket! A member of our team will reach out to you shortly.

recommeneded

json
    "transcriber": {
      "model": "nova-2",
      "language": "en",
      "provider": "deepgram",
      "endpointing": 300
    }

json
{
  "start_speaking_plan": {
    "waitSeconds": 0.8
  },
  "smartendpointing": true
}

@Karina Krinitskaya If you are certain about the duration you wish to allow for users to complete their speech when articulating numbers, you may continue utilizing the transcriptionEndpointingPlan. However, please refrain from making adjustments unless you are confident in your decision. Additionally, the issue regarding the numbers not being captured primarily is because of the low transcript confidence score and secondarily due to the endpointing plan. Therefore, I encourage you to implement the suggested changes and inform me of the outcome.